CHANGES OF MIND
An Essay on Rational Belief Revision
n e i l t e n na n t
PREFACE
This book sets off right away with the authors preferred account. Only thereafter
does it turn to the task of providing critiques of, or comparisons with, some of
the main opposing schools. This decision is occasioned by the wish not to fall
victim to what may be called the Chapter Syndrome. This syndrome is that the
majority of readers, for whatever reasons (skepticism? boredom? interruptions
beyond their control?) seldom progress past Chapter of any book they pick
up. One is tempted by the thought that this is why it is Chapter of Quines
celebrated Word and Object that gets all the attention. Whatever the reason for
the syndrome, this author is anxious to avoid it.
The positive account is therefore given in Part I. It is also given in what might
be called its natural logical order. The reader is introduced to the intuitive
motivations for, and the most natural and economical methods to apply in, the
process of rational belief revision.
The stress here on rational signals that it is a theory of competence that is
being proposed, not a descriptive theory of how ordinary thinkers do actually
go about changing their beliefs. The author is a logician, not a cognitive psychologist. Teaching introductory courses in logic has convinced him, sadly, that
even very intelligent people more often than not reason fallaciously. So the laws
of logic (or: rational deduction) do not describe how they actually reason. The
same people also have a hard time abandoning their cherished beliefs in the light
of recalcitrant evidence. So, whatever laws of rational belief revision one may
wish to propose, they are hardly likely to describe how even very intelligent
let alone, ordinarypeople actually adjust their beliefs in light of the evidence
available to them. What is on offer here, by contrast, is a rational account of how
one ought to go about changing ones beliefs. And at a more philosophically
P R E FAC E
vii
viii
P R E FAC E
ment.3 Tennant [b] set out the case for AGM-theorys collapse. Section .. of the present work provides a condensed account of the main results
of that paper. It is worth pointing out here that the word Degeneracy in its
titleA Degeneracy Theorem for the Full AGM-Theory of Theory-Revision
was being used in its familiar mathematical sense (and was even suggested to
the author by a mathematician colleague). That special sense involves trivial
or extremal casestypically ones that one wishes to avoid, in order to have
a good theory dealing with the non-trivial, non-extremal cases. The reader
could, however, justifiably associate with degeneracy the Lakatosian notion of
a degenerating research program. Rohit Parikh has called the result in question
a triviality result (for AGM-theory).4
This much having been said already about AGM-theory, it is worth saying
a little more in order to forestall any mistaken impression, on the part of the
reader, that the present work is perforce a commentary on, or sustained critique of, AGM-theory itself. It is not. The theory of rational belief revision on
offer here is developed within a completely different formal framework from
AGM-theory, and strikes out in a completely different direction from AGMtheory. This new way pays careful attention to epistemological considerations
(the structure of reasons underlying an agents beliefs). It represents belief systems as finitary, hence as potentially computationally tractable. It then provides
explicit algorithms for contracting and revising belief systems, thereby showing
that they are indeed computationally tractable. (Confining oneself to finitary
objects provides no guarantee that all the relations among them, and operations
on them, that might hold ones interest will be computable ones; but, in the
belief-revision case, they are.)
This work could have been written with no mention at all of AGM-theory.
For the account is self-contained. In deference to scholarly tradition and expectations, however, the author has been careful to make some central points of
connection and contrast between his own account and the main extant accounts
of other theorists in the three most relevant communities: epistemologists,
logicians, and computer scientists. AGM-theory is but one of the alternative
accounts considered in this connection.
Writing at treatise length on this topic has imposed severe interdisciplinary
demands. These include philosophical analysis, mathematical formalization,
See Churchs celebrated review of the second edition of Ayers Language, Truth and Logic, in
The Journal of Symbolic Logic, .
See Parikh [], p. , n. .
P R E FAC E
ix
P R E FAC E
ACKNOWLEDGMENTS
xi
a Fellowship lasting an academic year. These awards provided full relief from
teaching duties to enable pursuit of research for this book.
The author is grateful to colloquium audiences who have responded to presentations of his ideas on a computational theory of theory change: the Moral
Sciences Club at the University of Cambridge; the Workshop on Philosophy
of Physics at the University of St. Andrews; the Colloquium in History and
Philosophy of Science at the University of Colorado, Boulder; the th Annual
Workshop of the Center for Philosophy of Science in Pittsburgh; the Arch
Colloquium at the University of St. Andrews; the Dresden Workshop on Constructive Negation; the Princeton Philosophy Colloquium; the Central Division Meeting of the American Philosophical Association (where Joongol Kim
was the commentator); and, at Ohio State: the Research Group in Re-usable
Software; the Center for Cognitive Science; and an Interdisciplinary Research
Seminar on the Mechanization of Inference.
Much of the final draft was composed and revised in a little refuge in an
athletic facility at Ohio State, affectionately known by its users as the man cave.
All mens miseries come from one thing, which is not knowing how to sit quietly
in a room.5 It has been a salutary experience to be immersed in an atmosphere
of intense commitment on the part of coaches and elite athletes in Olympic
sports, and to sense the self-discipline, the focus and the dedication to hard
work that pervades the whole establishment.
Last, and most importantly, I want to express my love and gratitude to my
wife Janel Hall and our sons Henry and Hugo. They have put up with various
distractions and disruptions that the writing and proofing of this book have
occasioned.
tout le malheur des hommes vient dune seule chose, qui est de ne savoir pas demeurer en
repos, dans une chambre. Pascal, Penses, .
xii
AC K N OW L E D G M E N T S
CONTENTS
i n t ro d u c t i o n
. Logical statics v. logical dynamics
.. The fiction of the logical saint
.. Epistemic norms and the paragon
.. Classifying changes of mind
.
.
.
.
.
.
.
.
.. Applying the lessons learned in the case study from deductive logic
I Computational Considerations
c o m p u t i n g c ha n g e s i n b e l i e f
xiii
g l o ba l c o n d i t i o n s o n c o n t r ac t i o n
.
.
.
.
a f o r m a l t h e o ry o f c o n t r ac t i o n
. Introduction
. The computational complexity of various decision problems in
logic
. Relative entrenchment of beliefs
. Some intuitions
. The finitary predicament and the question of well-foundedness
. The logical macro-level v. micro-level
xiv
CONTENTS
. Closure
.. Universal closure of belief schemes
. Considerations of complexity
. Conclusion
s pe c i f i c at i o n o f a c o n t r ac t i o n a l g o ri t h m
.
.
.
.
Greedy algorithms
A brute-force, undiscerning algorithm
Preliminaries
Contraction Algorithm: First Version
.. Example of Version at work
.. An observation about the order of whitening
.. Discussion of Version
CONTENTS
xv
. Conclusion
a pro l o g pro g r a m f o r c o n t r ac t i o n
. Management of input files of problems
. The top-level part of the program
.. For search without checks for minimal mutilation
.. For search with checks for minimal mutilation
.. The lower-level part of the program
. Pre-processing
. Clauses for the various print commands
re s u lt s o f ru n n i n g o u r pro g r a m
f o r c o n t r ac t i o n
. Some simple contraction problems
. Outputs of our program on the foregoing problems
. The upshot
xvi
CONTENTS
t h e f i n i ta ry pre d i c a m e n t
m at h e m at i c a l j u s t i f i c at i o n s a re n o t
i n f i n i t e ly va ri o u s
. Definitions of terminology
. Results
III Comparisons
d i f f e re n c e s f ro m o t h e r f o r m a l t h e o ri e s
. Truth-maintenance systems
. AGM-theory
.. Some background
.. A brief summary of the AGM-theory of contraction
.. Implementability
.. The Postulate of Recovery
.. Minimal mutilation
.. The degeneracy of AGM-theory
.. A condensed account of the degeneracy results
. Bayesian Networks
c o n n e c t i o n s w i t h va ri o u s e pi s t e m o l o g i c a l
ac c o u n t s
. Stage setting
. Perennial topics of mainstream epistemology
.. Skepticism
.. Basic beliefs
.. Foundationalism, coherentism and foundherentism
CONTENTS
xvii
References
Index
xviii
CONTENTS
CHAPTER 1
Introduction
his is a book on the dynamics of theory changeor, as it is now commonly called, belief contraction and belief revision. It provides a new
picture of how we ought to change our minds when the need arises. It
begins with that picture, and then examines the main current competitors, in
order to explain the comparative advantages of the new account. The book is
therefore an invitation to those other theorists to engage in an exercise of the
very method that we think about.
Ever since Hume, philosophers have been aware of the need to assign equal
importance to the propositional attitudes of belief and desirethe one arising
from perception, the other leading to action. As John Searle puts it, belief
involves the mind-to-world fit, whereas desire involves the world-to-mind fit.1
And in Donald Davidsons view of the mind, reasons for an agents intentional
actions are belief-desire pairs.2 Hence, contemporary philosophers might well
take changes of mind to encompass changes in desire as well as changes in
belief. The title of this book, however, is a little more restrictive. Changes in
desires the author would call changes of heart; so that changes of mind are
changes of belief. The reader will, one trusts, permit this quaint distinction. The
discussion here will concern only changes in belief. If that leaves out half the
Searle first introduced the contrast in Searle [], at p. . See also Searle [], pp. ,
and Searle [], pp. .
See Davidson [], p. .
INTRODUCTION
picture, it does not really matter. The author has nothing systematic, rational or
useful to say about changing ones desiresand if he did, he would be unlikely
to share it.
INTRODUCTION
lifetimes. So, even if they only ever reason impeccably, never making any logical
mistakes, their belief systems could never be theories in the logicians sense.
In order to make vivid the difference between the statics and the dynamics
of belief, however, let us return for a moment to the fiction of the logical saint.
Think of a logical saints current beliefs as forming a logically closed set in the
logically specious present. Even a logical saint, however, would have to take at
least one discrete quantum of time, no matter how short, in order to change its
mind. Otherwise there would be no way to distinguish between what the saint
formerly believed, and what she now believes. For the logical saint, change still
takes time; whereas logical closure does not.
One also wishes to succeed in making whatever change one has decided upon.
Suppose the change in question is to surrender a certain belief. Then one would
not wish to have that belief popping back up in the next logically specious
present. This could happen if there were enough doxastic detritus left behind
after the initial attempted excision. Surrendering beliefs involves undertaking a
thorough purging. One must not leave behind other beliefs that would rationally
demand continued commitment to the belief to be surrendered.
Sometimes a change of terminology is needed in order to slough off old mental habits. So the talk here is of surrendering and switching beliefs, rather than
(or in addition to) talking of belief contraction and belief revision. The reason is
this: belief contraction is something of a solecism; it should, more accurately,
be written as contraction of a system of beliefs. It is the system that contracts
when one gives up a belief. It does not make sense to think of an individual belief
being contracted. (There are no font-size reductions in Mentalese ) Likewise,
we may revise a system of beliefs by following through on an initial decision to
switch from belief in a certain proposition to disbelief. That is, one can switch
from the belief that p to the belief that p; and in so doing, one would in general
need to make yet further changes, changes that amount to a successful revision
of ones system of beliefs.
The already expert reader must be careful not to invest the terms contraction
and revision, which are intended to apply here to systems of belief, with the narrowly technical meanings that may be bestowed on them by current paradigms
of theory change. An altogether different account will be offered here of the
content and structure of a system of beliefs, and of the systematic changes that
we call contraction and revision. Unless and until we are actually discussing
competing accounts of contraction and revision, the reader must be prepared to
take these terms at this stage as pre-formal, and in need of eventual explication.
incremental change
A metabelief that has gained wide currency from the work of Thomas Kuhn
on scientific revolutions is that people change their beliefs somewhat in the
way that moths changed their coloration: by having old beliefs die out with
their bearers, and having new beliefs take hold in the next generation. One
mentions moths here because the phenomenon of industrial melanism is one
of the standard examples used in order to illustrate the workings of natu4
INTRODUCTION
ral selection. Unlike chameleons, individual moths cannot change their coloration during their lifetimes. But coloration can vary within any population
of moths: some will be lighter than average, others darker. During the industrial revolution, the dark satanic mills belched out fumes that darkened surrounding tree trunks. So lighter-colored moths were easier pickings for birds.
Darker-colored moths were therefore at a selective advantage. Within a few
generations, the proportion of darker-colored moths rose significantly. Moths
changed their coloration because the lighter-colored ones died out, leaving
relatively fewer offspring like them, and darker coloration took hold in the next
generation.
Consider now the following flamboyant and Kuhnian comment on the dead
weight of dogma (from Bayles and Orland [], at p. ):
When Columbus returned from the New World and proclaimed the earth was round,
almost everyone else went right on believing the earth was flat. Then they diedand the
next generation grew up believing the world was round. Thats how people change their
minds.
One is reminded also of the dark quip that social progress (which involves the
adoption of more enlightened beliefs) takes place one funeral at a time.
But for the more rational among us, that is not how we change our minds;
or, at least, that is not how we ought to change our minds. Indeed, this picture
of the dead weight of dogma does not make any place at all for people actually
changing their mindsany more than the evolutionary account of industrial
melanism represents individual moths as changing their coloration during their
own lifetimes.
What we want is a more chameleonic account of rational belief change. We
need to show how individual agents can and should change their own beliefs
whether as the result of perception, or the testimony of others, or theoretical
re-consideration. Consider the following examples of how individuals might
change their minds.
John has called his engagement off. For he has just discovered that his fiance
Mary has a dark secret. The discovery conflicts with many of his previously
cherished beliefs about her. He no longer believes she is trustworthy. Nor does
he believe that they will marry. But should he change any of his beliefs about the
institution of marriage? Should he call into question his belief that he will, one
day, be able to find the right partner? Should he revisit his former beliefand
now, perhaps, view it as navely trustingthat there is no such thing as a battle
between the sexes? How far should his belief revision go?
R E VO L U T I O N A RY C H A N G E V. N O R M A L C H A N G E
INTRODUCTION
she will not seek to make contribute any would-be justificatory support to
any other beliefs she might form. For, as she knows, logical truths can always
be suppressed as premises. So, in the absence of any conflicts among beliefs,
changes of mind on the part of a logical paragon can be presumed to involve
only propositions that, as far as the paragon knows, are logically contingent. (She
can, of course, hold mathematical beliefs, on the basis of logical proof from
mathematical axioms; but these will not be logical truths in the conventional
sense, unless some extremely implausible logicist account of mathematics turns
out to be true.)
It is worth dwelling further on the foregoing emphasis on the rider as far as
the paragon knows. Unlike a saint, a paragon is not presumed to be logically
omniscient. She is always right about the logical proofs she constructs, and she
remembers them; but she is not presumed to be aware of every logical deducibility, except for those that hold in effectively decidable systems of logic. (And
even then, she would need to be given the time it would take to reach a correct
decision on the logical validity of any argument that she might be required to
appraise. Her knowledge of such deducibilities would not be instantaneous, as
is presumed to be the case with a logical saint.) The following situation is therefore possible, for a paragon. She subscribes to classical first-order logic, whose
decision problem is not effectively decidable. Two contingent and logically complex propositions p and p , let us say, are (classically) logically equivalent, but
the paragon does not realize this, because of the effective undecidability just
mentioned. The paragon comes to believe that p. By a step of -Introduction,
she then comes to believe that p p . She does not realize that the latter
is a (classical) logical truth. Subsequently, she may decide to surrender this
disjunctive belief. So, her change of mind, in this example, happens to involve
a logical truth. Still, as far as the paragon knows, the disjunctive belief p p
could be contingent.
If an objector wishes to avoid such hair-splitting cases, we could simply stipulate, at the outset, that our formal modeling is confined to belief schemes that
consist only of genuinely logically contingent propositionseven if the rational
agent in question happens to be unaware that they are all indeed logically
contingent, and lacks any effective method for establishing that this is indeed
the case. Such stipulation would be unnecessary, however, for those cases where
theoremhood in the paragons logic is effectively decidable.
Our omission of logically true sentences (as premises or conclusions of steps)
can be justified further as follows. All theorists of theory contraction agree that
the contraction of T with respect to any logical truth is simply T. Put another
8
INTRODUCTION
way, one cannot really contract with respect to any logical truth. Hence, there
is also no point at all in insisting on logical truths being able to appear as
conclusions of any steps of systems that are eligible for contraction with respect
to any of their nodes. Moreover, it is a well-known point of logic that a logical
truth can always be suppressed wherever it may appear as a premise in an
argument. Hence, there is no point either in insisting on logical truths being
able to appear as premises of any steps. We realize, of coursesee above
that a logical paragon, according to our definition of this idealized kind of
agent, might be using a logic with an effectively undecidable decision problem.
Therefore we cannot assume that she would at any given time know, of any
logical truth in her language, that it is indeed a logical truth. This means that she
might occasionally have, in her belief network, a node labeled by a sentence of
her language that happens to be a logical truth, like the p p example above.
Should she eventually discover that it has this status, however, she will simply
omit the node, along with all and any justificatory connections it had enjoyed,
from her belief network. Getting the logical status of any particular node wrong
is no principled obstacle to our modeling a rational agents belief nodes as
though they were labeled only by contingent sentences. For the contraction
process, by her own lights, would, at any time, still have to proceed according
to the rules we are laying down.
The ideal of logical saintliness is an extreme one. The logical saint is required
to just know that a given proposition follows logically from certain others when
indeed it does. Alonzo Churchs undecidability theorem for first-order logic tells
us that it is in principle impossible for any mechanically simulable intelligence
to decide, of any given argument P , . . . , Pn , ergo Q, and within a finite period
of time, whether there is a logical proof of the conclusion Q from the premises
P , . . . , Pn . So any saintly logician who can do that (for all P , . . . , Pn , Q) is
beyond the limits of mechanical simulation. And that is the case even absent
the further requirement that the knowledge should be available instantaneously
whenever the saint needs it.
We are proposing a lesser extreme, however, which can still be useful in
guiding our endeavors. We have weakened the requirements for good logical conduct. Suppose we hold the agent responsible only for never neglecting
those logical transitions among sentences for which she has already worked
out a justification. Just as she is allowed to be perceptually mistaken, from
time to time, so too now she can be allowed to be ignorant of certain logical
consequences. But she is required never to forget and never to fail to apply,
where appropriate, any justificatory transition for which she has already worked
M O R E O N PA R AG O N S A N D S A I N T S
out satisfactory details. Moreover, we require that, within the limitations of her
explicit resources, any mechanically decidable logical matter is one on which
she will make a correct decision (given enough time). We are calling this more
modest kind of norm-obeying agent a paragon, or a rational agent. Rational
agents are, after all, resource bounded. We require them only to work perfectly
with what theyve got. And we must allow them whatever time it necessarily takes
for them to do so.
To aspire to saintliness is to aspire for the impossible. To aim to be a paragon,
however, is a realistic ideal. This book is about how a paragon would change its
beliefs.
proposition
If the state of mind A changes to the different state of mind B, we shall write
A B. (This is not to be confused with the conditional if A then B.) There are,
in all, six kinds of initial change that could occur, concerning any contingent
proposition p :
1. believe p neither believe nor disbelieve p;
2. believe p disbelieve p;
3. neither believe nor disbelieve p believe p;
4. neither believe nor disbelieve p disbelieve p;
5. disbelieve p believe p;
6. disbelieve p neither believe nor disbelieve p.
In the classical case, when one is using classical logic, p is logically equivalent to
p. So () can be re-written as
10
INTRODUCTION
5 . believe p believe p,
upon which we see that both () and ( ) have the common form
Switching: believe believe .
Likewise, () and () have the common form
Surrendering: believe neither believe nor believe .
Finally, () and () have the common form
Adopting: neither believe nor believe believe .
When switching a belief, one revises ones system of belief. When surrendering
a belief, one contracts ones system of belief. And when adopting a belief, one
expands ones system of belief.
11
12
INTRODUCTION
of finitary objects and structures. This means that we can also inquire after the
implementation of those norms in a computational system.
13
14
INTRODUCTION
See Section ..
C O M P U TAT I O N A L C O N S I D E R AT I O N S
15
16
INTRODUCTION
for the agent, by virtue of instantiating a valid inference within the agents
chosen logic, such steps would cease to be justificatoryand would need to
be surrendered, or given upif the agent were to revise her logic in such a way
as to invalidate the formerly valid inference pattern in question. (The inference
from xF(x) to xF(x), for example, would be invalidated in the process of
revising classical logic so as to yield intuitionistic logic.)
Such an extension of our account, however, is not undertaken here. The
ensuing discussion will be conducted on the assumption that the justificatory
steps are immune to retraction or revision, and that it is only the beliefs at
their premise and/or conclusion nodes, so to speak, that can be adopted or
surrendered. But it is important to keep an open mind about the possibility
of rendering such steps liable to retraction, in situations where it is not enough
to change the doxastic status only of the nodes that they connect. Having said
that, it will emerge that one rationally cannot give up any justificatory step that
is an instance of a valid argument in the Core Logic just mentioned. For that
would lead to an overall reflexive instability in the mental makeup of the agent.
She would no longer be able to conduct the reasoning that is required in order
to motivate the very project of belief revision that is rationally called for when
one discovers that ones current belief scheme is contradicted by newly available
evidence. If one gives up any of the rules of Core Logic, then one will not be in
a position to work out what to do when placed in such a predicament. Such
an unstable situation would not be irredeemablethe agent would just have to
take back on board (at least) the inference rules of Core Logic.
The behavior of the new contraction algorithm deserves to be investigated
systematically on many larger problem sets, in search of regularities involving the initial structure of a belief system and the variety of ways in which
one can contract it. Of particular interest will be phase-transition or threshold phenomena such as those encountered at times of theoretical crisis, when
many anomalies have cropped up. Thomas Kuhn wrote of paradigm shifts,
or revolutionary theory changesuch as the shift from Newtonian dynamics
to Einsteinian relativity theory. What prepares the ground for such a shift?
Why does the old theory suddenly implode under the force of the contractions demanded by experience? How does contraction in response to bits of
evidence induce large-scale theoretical collapse? Might there be a more deeply
rational process at work than some of Kuhns followers have claimed? The
work seeks to provide new methods that could throw some light on these
questions.
P H I L O S O P H I C A L C O N S I D E R AT I O N S
17
1.8 Methods
As mentioned in the Preface, this interdisciplinary project requires philosophical analysis, mathematical formalization, logical systematization, and application of the main procedures in computing science: choice of data types (also
sometimes called data structures); isolation of effectively decidable properties
and relations; analysis of computational complexity; specification of algorithms;
and their implementation in high-level computer languages. Contraction algorithms will vary in their optimality and efficiency, in light of NP-completeness.
They are specified in detail, so that Artificial Intelligence (AI)-theorists can
implement them in various computing languages. A Prolog program is also
provided for the simplest version of the contraction algorithm.
Paul Thagard was one of the first philosophers of science to emphasize the
importance of AI and computation for our understanding of scientific theories. In Thagard [] he described how an Artificial Intelligence (AI)-based
approach could throw light on the discovery, testing and evaluation of scientific
theories. This involved consideration of analogical problem solving, abduction,
and inference to the best explanation. Thagards preference was for very rich
data types, and programs that worked well by delving into, exploring, and
making connections among the innards of these data types. He chose formalizations that were easy to encode in the programming language LISP, for
such things as concepts, rules, problems, and problem-schemata. As he stressed
(pp. ), the
structures and algorithms are interdependent: the algorithms must work with the data in
the form given to them.Philosophers tend to assume the ubiquity of only one kind of
data structurethe propositionand only one kind of algorithmlogical reasoning. But
computer science offers a wealth of structures in which data can be stored: arrays, tables,
records, frames, and so on.[O]ur view of the nature of thinking can be broadened considerably by allowing for the possibility of nonpropositional data structures and nonlogistic
processing mechanisms.
Here we follow Thagards advice, but do not follow his example. We definitely get
away from the proposition as the, or even the main, data type. Instead, our only
data type of any significance is the step, consisting of finitely many premises and
a conclusion. But the way in which we do not follow Thagards own example is
that the nodes of which our steps consist are primitive Urelemente, and internally unstructured. And apart from defining dependency networks as finite
sets of steps satisfying certain very simple conditions, that is it; that is all we
deal with.
18
INTRODUCTION
M E T H O D O L O G I C A L C O N S I D E R AT I O N S
19
basic virtues of a formal model. Cautionary note for logicians: here we are using
model in the informal sense of scientific model (i.e. theory) of some subject
matter, rather than in the specialized sense in which logicians talk of models of
theories in a formal language. Models in the latter sense are perhaps models, in
the former sense, of the kinds of subject matter that such languages can describe.
SIMPLICITY
Consider the classical propositional calculus. It is a simple model of deductive reasoningindeed, the simplest possible. (Sometimes experts call it baby
logic; this underscores the point being made here.)
The propositional calculus deals only with the logical connectivesnegation,
conjunction, disjunction and the conditional. The formal semantics for a propositional language tells one how to interpret its atomic sentences (the so-called
propositional variables) by assigning truth values to themtrue or false. And the
semantics interprets the connectives by means of the well-known truth tables,
which tell one how to compute the truth value of a complex sentence, under
an assignment of truth values to its atomic constituents. The notion of logical
consequence can then be defined as preservation of the truth value true from
premises to conclusion, under all possible assignments:
is a logical consequence of 1 , . . . , n
if and only if
for every assignment of truth values to atomic sentences occurring therein, if makes
all of 1 , . . . , n true, then makes true.
The reader who is not a logician might wish to take a quick look at the formal rules for Core
Logic, which are stated in Section ..
20
INTRODUCTION
This brief explanation of the theory of propositional logic will have to suffice,
for it is provided only in order to illustrate our methodological point about the
need to strike a balance between the demand for simplicity and the demand for
comprehensiveness.
The simplicity of propositional logic is self-evident. Against this setting
one attains a very good and serviceable conceptual grasp of the distinction
between syntax and semantics, and of the metalogical notions of soundness
and completeness. One could go further and develop computational methods
of proof search within this system, but this is not usually done in standard
introductory texts. 15 One could also illustrate various metalogical themes
arising from the study of fragments of the language, such as the fragment based
on just negation and conjunction. 16 One could study certain preferred forms
of proof that exhibit a particular sort of elegance and directness. These are the
proofs in so-called normal form. (Again, this is a topic not usually covered in
standard introductory texts.)
The point emerging from these remarks is this: simple though it is, propositional logicwhen presented in the right wayis a wonderful setting within
See the discussion in Section . of the complexity of the decision problems for various wellknown systems of propositional logic.
A good methodological dictum about logical theorizing, acquired by this author from his
teacher Timothy Smiley, is that if ever you wish to establish a metalogical result, check first how it
works out for just negation and conjunction!
M E T H O D O L O G I C A L C O N S I D E R AT I O N S
21
which to get to grips with, and to learn standard theoretical techniques for
proving, results that can be extended to more complicated systems containing all
of propositional logic as a subsystem. (We shall look at one of these presently.)
It is also worth remarking that even at the simple level of propositional
language and logic, a host of philosophical issues and debates can arise. These
concern such questions as:
What are the truth bearers? Are they sentences, or statements, or propositions?
Why do we assume that the truth value of any one atomic sentence can be
assigned to it independently of the truth values assigned to other atomic
sentences? Why do we restrict ourselves to single-conclusion arguments? Why
not study multiple-conclusion calculi? What is the right account of logical
consequence? Should we say that a logically contradictory set of premises such
as {A, A} has any and every sentence B as a logical consequence? Or should
we deny this (as the relevantist does)? 17 And what about determinacy of truth
value? Should we be assuming that every sentence is determinately true or false?
Or might there be truth-value gaps? . . . or even gluts? Is it really the case (as
the classicist maintains), that using only negation and a binary connective, one
can define all other connectives? What if we think (as the intuitionist does)
that each of the usual connectives (negation, conjunction, disjunction and the
conditional) cannot be defined in terms of the other three? And what about
alternative truth values, such as indeterminate (for quantum mechanics), or
evaluation transcendent (for deeply elusive conjectures in mathematics, such
as the Continuum Hypothesis), or in-between (for vague statements)? Indeed,
what about having a continuum of truth values, as is done in probability theory,
which uses the real unit interval [,] as its space of truth values?
The classical system of logic involves taking some peremptory (and in some
cases carefully considered) stands on these issues.
Truth bearers are sentences. Atomic sentences are semantically independent
of each other. We study single-conclusion calculi for the sake of simplicity,
but we could accommodate multiple-conclusion calculi if you wish. The right
account of logical consequence is preservation of classically construed truth.
Contradictions logically imply every sentence. Truth values are determinate, even
if their determination transcends our grasp. There cannot be truth-value gaps or
The inference from A and A to (arbitrary) B is known as Lewiss First Paradox. It is taken
as correct in intuitionistic, hence also in classical, logic. So both these systems fail to be systems of
relevant logic.
22
INTRODUCTION
gluts. Negation and one binary connective yield all other connectives by suitable
definitions. And aroint thee, knave, we shall have no truck with alternative truth
values!
Note that to get bogged down with the first question, concerning the right
choice of truth bearers, is to opt for philosophical disquisition while sacrificing even the simplest form of system building. If one were to become too
obsessed with finding philosophically satisfying answers to all the questions
initially raised, one might never get round to building any sort of deductive
system.
The system-building theorist, however, gets a headstart by being willing to
make certain idealizing assumptions, plumping for structures and methods
that, it is to be hoped, will prove to be invariant across different resolutions
of the philosophical issues raised. One can study the structure of deductive
inference without having to settle in advance the minutiae of the debate over
truth bearers. With other topics on the list above, however, one might have to
proceed more cautiously. On the question of relevance (does a contradiction
logically imply every sentence/statement/proposition?) one might need a better
idea of the shape of ones answer before deciding on any particular choice of
rules of inference. This is because those very rules are the ones that will be
generating proofs of any conclusions that do follow from premises that happen
to be contradictory.
Alternatively, on questions such as this one, which have serious implications
for ones choice of formal logic, one can be a system builder in advance of any
final resolution of the philosophical issue concerned. One can try to cater for
the various choices that could eventually represent informed answers to the
question in hand. Indeed, ones attempts in this connection at system building
might furnish valuable considerations that can be brought to bear in offering a
final answer to the question posed.
There would appear, then, to be two strategies open to the system builder,
when faced with a thorny philosophical issue:
1. try to furnish a system that is invariant across (hence: indifferent to) the various
possible resolutions of the issue; or
2. try to tailor different systems to the different possible outcomes associated with
competing resolutions of the issue.
23
24
INTRODUCTION
25
be made to look like the obvious leg-up that it is, in order to be able to deal
with the quantifiers as well as the connectives.
Having thus specified the slightly more complicated (because more comprehensive) formal semantics and system of proof for the quantifiers and the
identity predicate, one then finds that the work has already been done, at the
propositional level, in defining soundness and completeness. The very same definitions work at first order as they did before, provided only that one appropriately re-construe the formal proof system S and the formal language L therein.
For S now contains the inference rules governing the quantifiers and the identity predicate, and these expressions have increased the expressive power of L,
and have been characterized semantically.
The lift to first order brings some rather important theoretical changes. Most
conspicuous among these is that there is no longer any mechanical decision
procedure for telling whether there is a proof of a given conclusion from finitely
many given premises. And of course one requires a rather more sophisticated
method of proof, in the metalanguage, in order to establish that the formal firstorder proof system is complete. 19
26
INTRODUCTION
handled via nodes within our networks just as easily as any (linguistically) simpler beliefs, expressible by using more modest linguistic resources. In speaking
of our account of belief change as analogous to the logicians account of propositional logic, we are adverting, rather, to the systemic simplicity of propositional
logic itself.
We shall enter several idealizing assumptions, which are bound to provoke
debate:
Belief systems can be represented as finite dependency networks. Structureless
nodes represent the agents beliefs. Nodes are either in or out (believed or not
believed). Steps connecting the nodes represent apodeictic warrant-preserving
transitions (for the agent). Steps are either initial or transitional. Belief systems
are made up of steps.
We do not require our dependency networks to be structured in any particular
way. In particular, we do not insist that they form tree-like structures (though
of course they may). In a contraction process, steps cannot be surrendered; only
nodes can.
Our interest in making all these idealizing assumptions is that they will reduce
a thorny subject to more manageable proportions at the outset. They will allow
us to investigate some rather intriguing phenomena in the contraction and
revision of belief systems. They will afford insight into the computational complexity of the contraction process. They will enable a manageable first stab at
an actual implementation of a contraction algorithm, which runs efficiently in
real time on reasonably complex examples. This affords the prospect of further
theoretical insights into such matters as Kuhnian sudden collapse under the
weight of accumulating anomalies.
Moreover, the formal theory of contraction that we shall develop, and the
algorithms that we shall specify, and our Prolog implementations thereof, have
the desirable feature of not taking a doctrinaire stand on the major epistemological issue of foundationalism versus coherentism. But our modeling will offer the
intriguing possibility that, as one computationally explores the consequences,
for belief contraction and belief revision, of embodying (say) coherentist patterns of interdependence in belief schemes, one may learn crucial lessons about
the behavior of such schemes under contraction and revision. These lessons may
reflect well or poorly on the philosophical credentials claimed for coherentism.
There is nothing more sobering than certain results of formal exploration of
grand philosophical commitments: ask the ghost of Gottlob Frege about nave
abstraction, or the ghost of Alfred Ayer about indirect verifiability.
M E T H O D O L O G I C A L C O N S I D E R AT I O N S
27
The austere formal modeling that we develop here can serve as a departure
point for further investigations that might yield refinements and extensions of
the modest initial account. We shall touch on various possibilities in Section .,
so it is not necessary to go into details at this point.
28
INTRODUCTION
process of adjustment. That is, in giving up your belief p, you also give up q. The
Recovery Postulate claims that if you have an immediate change of mind, and
adopt once more the belief p, then you will thereby reclaim the belief q.
Such has been the view, until very recently, of AGM-theorists of belief contraction. They treat systems of belief as idealized, logically closed theories, and
study how one such theory can be a contraction, or revision, of another, with
respect to any particular sentence p. In other words, they think of the rational
agent as a logical saint in the sense explained above. A whole mathematical and
metalogical edifice, with representation theorems and exaggerated promises
of computational applications, was built upon this unexamined dogma about
the surrendering and re-capturing of beliefs. The dogma is false, however, and
ought itself to be surrendered. This book shows what can be recovered without
the Recovery Postulate.
Any new account should seek to remedy these shortcomings, and achieve further important goals. It should justify the claim that representing belief systems
as finite dependency networks (as a computational account must do) incurs no
loss of generality, and no restriction in the scope of applicability. It should establish how complex the contraction problem is. It should also explicate the notion
of minimal mutilation. By working on the way we employ justificatory relations
among propositions when contracting a belief system, we aim to uncover the
essential features of theory change in general, without resorting to non-standard
revisions of the underlying logic itself.
summary
. Introduction (i.e. this chapter)
We have explained the distinction between, on the one hand, logic as a theory of belief
statics and, on the other hand, our sought account of belief dynamics. The various kinds
of belief change have been classified. These are: surrendering, adopting or switching individual beliefs; and thereby contracting, expanding or revising ones system of beliefs. Our
account of the epistemic norms involved is agent-centric. The idealized figure of the logical paragon (as opposed to the completely fictional figure of the logical saint) is introduced as the guiding model of a rational agent who is thoroughly competent in matters of
belief change. We discuss what a theory of belief change needs to characterize or make
feasible. Two key constraints are formulated: we need to explicate (and ensure) both
minimal mutilation and minimal bloating of systems of belief undergoing contractions
C H A P T E R - B Y- C H A P T E R F O R E S H A D OW I N G S U M M A RY
29
30
INTRODUCTION
We work through many small examples to impart a thorough and vivid understanding
of the dynamics of belief change, using these conventions.
C H A P T E R - B Y- C H A P T E R F O R E S H A D OW I N G S U M M A RY
31
32
INTRODUCTION
C H A P T E R - B Y- C H A P T E R F O R E S H A D OW I N G S U M M A RY
33
PA RT I
Computational Considerations
CHAPTER 2
Computing Changes
in Belief
37
that the epistemological explanations of our modeling are offered only in order
to motivate our formal concepts. The formal material itself, however, when
considered in its own right, independently of the epistemological interpretations, is mathematically (or combinatorially) concrete enough to allow one to
pose, and to solve, questions about the computational complexity of contraction
problems.
A dependency network is a useful formal representation of a rational agents
belief scheme. We shall now introduce local features of dependency networks,
for which we shall provide diagrams. Our diagrams are designed to represent
sentences believed (black), sentences not believed (white), and justificatory
steps involving such sentences, mediated by inference strokes. The idea of
including inference strokes as integral parts of a diagram of dependencies (of
conclusions on sets of premises) is taken from Chapter , Graph Arguments,
of the pioneering work of Shoesmith and Smiley [] in multiple-conclusion
logic. Even though we are not here contemplating steps with multiple conclusions (see below), the use of inference strokes within our diagrams affords a
structural clarity that has proved to be indispensable in our dealing with the
problems of contraction and revision of belief schemes.
38
just mentioned. We need to have an eye not only on the nodes that are involved
but also on the steps that connect them.
We need, henceforth, to be relieved of the obligation to keep putting in the
parenthetical riders (for the agent in question). So the reader is put on notice
that all of our modeling is thoroughly agent-centric. That said, we shall still insert
an occasional reminder to this effect.
Note that not believing a proposition p does not entail believing its negation
p, i.e. disbelieving p. A proposition that is not believed can be moot or undecidedneither believed nor disbelieved.
We shall henceforth permit ourselves the solecism believing a sentence.
Black means in and white means out, as far as the belief scheme is concerned.
We do not need to distinguish between sentences that are neither believed nor
disbelieved, and sentences that are disbelieved. Both these kinds of sentence are
considered out, and would be rendered white.
With this initial step we have a black inference stroke sending out an arrow to
the black node labelled a. No arrows come into the inference stroke in question;
and this indicates that a (from the agents point of view) may be, and is, believed
outright. Call a stroke that receives no arrows a starting stroke.
Epistemologists are divided on the question of what kind of belief may serve
as an initial belief for a rational agent, especially if one requires the agents
belief scheme to exhibit a foundationalist structure. The safest and least controversial examplesat least for one whose commonsense intuitions survive
exposure to the problem of skepticismare simple observation reports, such
as that is a brown penny. As the old adage has it, seeing is believingeven if,
in response to the skeptic, one needs to emphasize how such simple knowledge
claims can be corroborated by bringing to bear other sensory modalities, such as
touch.
39
. . . bn
carries the logical interpretation (for an agent who adopts the step)
if one is justified in believing each of b1 , . . . , bn ,
then one is justified in believing a.
. . . bn
With this transitional step we have the black inference stroke receiving arrows
from (exactly) the black nodes labeled b , . . . , bn , and sending out an arrow
to the black node labeled a. It is by means of its inference stroke that a step can
be identified. For, given the inference stroke, one determines the premises of the
40
step as those nodes that send arrows to the inference stroke; and one determines
the conclusion of the step as the sole node to which the inference stroke sends
an arrow.
It may come as a surprise to the reader to learn that a node can be the
conclusion both of an initial step and of a transitional stepwe shall have more
to say about this feature in Section ...
Transitional steps are not all confined to the agents belief box. Put in terms
of our chosen coloration convention: it is not necessary that all transitional
steps known to the agent should have black inference strokes; some of them
could be white. The agent could be apprised of a deductive transition, say, from
b , . . . , bn to a, without believing a. This would entail that the agent (whom we
are assuming to be rational) would not believe all of b , . . . , bn . Suppose for
the sake of illustration that she does not believe either a or b but does believe
b , . . . , bn . Such a step would be rendered in black and white as follows:
b1 b2
...
bn
Note that the inference stroke is white, as indicated by the white space created
between the two thin parallel lines in place of what would otherwise be a
single inference stroke. This notation is suggestive also of a decoupling of the
premises from the conclusion, showing that the latter does not acquire support
from the former (collectively), since not all of them are black. The whitening
of the inference stroke does not mean that the agent has come to doubt the
inference concerned. Rather, it means that this inference is not responsible for
the agents belief in the conclusion a (should she indeed happen to believe a).
For, even though the step is still (as far as the agent is concerned) justificatorily
valid, it fails to confer belief on a because not all its premises are believed. This
would continue to be the case even if a were believed, but on the basis, say, of
other beliefs c , . . . , cm . In such a case the previous diagram would continue to
contribute its white inference stroke, but the node for a would be black, on the
strength of the black inference stroke transmitting the support of c , . . . , cm , all
of which are black:
41
b1 b2
...
bn
c1
...
cm
The two premise sets could, in general, have members in common. All that is
required, for any node a, is that each of two distinct premise sets of steps with
conclusion a should have a member that is not in the other premise set. That
entails, in particular, that neither set may be included in the other (more on this
below). We can call two such sets non-inclusively distinct.
In this and subsequent chapters, such an inference (with n ) will be taken
to enjoy an inviolable status, for the agent in question. Unlike initial beliefs (for
which n = ), such an inference is not one that an agent will be able to give
up, once she has adopted it (once she concurs with the transmission of justification that it represents for her). This means that the step (from {b , . . . , bn }
to a, where n ) remains in the agents inferential base forever, once she has
acquired it. Unlike its propositional nodes, such a step cannot be given up. The
agent might surrender her belief in a, and therefore also in at least one of the
premises b , . . . , bn ; she would not, however, expunge the step itself from her
accumulated base of such steps. (It is worth reminding the reader that this
means that we cannot, at this stage of our investigations, offer a formal account
of any process of revising ones logic. But the basic account could eventually be
extended so as to deal with such revision.)
42
This distinction that we are proposing to make between a step and its nodes
(as to their revisability for the agent) is a methodological one, with the aim of
simplifying the modeling for the time being. We maintain it for as long as we
can, in our attempt to model the process of rational contraction and revision
of belief schemes. One can think of the set of steps adopted by a given agent
as forming a sort of jungle gym of fixed connections among nodes. The nodes
that are believed will be colored black; all other nodes will be colored white. It
is possible that at any given time many nodes are white.
Contractions and revisions involve changes in patterns of coloration only.
They do not involve adding to, or destroying any part of, the agents jungle
gym of justificatory steps (her inferential basis). In terminology to be introduced below, the structure of nodes and of steps connecting them remains
constant, while the coloration of the nodes (and of the steps inference strokes)
may change.
It lies beyond the scope of this work to inquire into the possibility of making
steps themselves eligible for expulsion or elimination from an inferential basis.
This will probably bring with it the prospect of drastic changes taking place
within a belief scheme. Thus, many steps in a single agents scheme can be
there by virtue of a common form that the agent has descried. (Instances of
inductive or certain deductive rules of inference serve as an example here.)
Presumably, the expulsion of any one step of such a form would require the
expulsion of all the other steps of that same form. For, if each of them is in
the agents scheme by virtue of its having a certain form, what could rationalize
holding on to any one of them, once one of them has been given up? If we make
a particular transition in thought only on the basis of its form, we always blame
the form if the transition is found to be in error. And we subsequently mistrust
all transitions of that form, whose correctness was once supposed to consist in
their having the form in question. These considerations reveal both the power
and the potential frailty of formal transitions. The revealed defectiveness of any
one instance of that form discredits the form itself as a basis for confidence in
any other of its instances.
Of course, any step {b , . . . , bn }, therefore a:
b1
. . . bn
43
can be rendered absolutely invulnerable (for the agent) by encoding the transition it represents as the extra premise (b . . . bn ) a, thereby ensuring
that the resulting (n+)-premise step is logically necessary:
b1
. . . bn
(b1 . . . bn) a
What this means is that the agent has packed into the sentential belief
(b . . . bn ) a
the burden of justificatory transmission of the erstwhile n-premise step. The
new step with (n+) premises, being core-logically valid, will be inviolable for
any rational agent. Core logic is a very special subsystem of intuitionistic (hence
of classical) logic. We shall be arguing that it is the part of logic that cannot
rationally be revised, if we wish to maintain a certain reflexive stability for the
whole enterprise of rational belief revision. We shall be returning to this theme
in Chapter . The core proof of the (n+)-premise step involves multiple steps
of -Introduction (to take one from b , . . . , bn to b . . . bn ) followed by
a step of -Elimination with (b . . . bn ) a as its major premise. Both
-Introduction and -Elimination are rules of Core Logic.
We adopt the rational simplification that if the step from {b , . . . , bn } to a is
known to the agent, then she will not know of any step to a from any proper subset of {b , . . . , bn }. This is because the agent, being rational, will be concerned to
remember only steps with premise sets that are (as far as she can tell at the time)
inclusion-minimal in justifying the conclusion. Why should any rational agent,
after all, bother to remember any dilution, or weakening,1 of a step of which
she is already apprised? If she already knows that a follows from b alone, why
should she bother to bear in mind the diluting step from {b , b } to a? In the
interests of efficiency, and economical use of resources, she should undertake to
remember only the stronger step, which uses the reduced set {b } of premises.
The foregoing rational simplification does not ensure, however, that every
step {b , . . . , bn }, therefore a known to the agent really needs all its premises
44
Logicians will recognize this borrowing of terminology from Gentzens sequent calculi.
in order to be valid, or justificatory (in the appropriately objective, agentindependent sense that attaches, say, to logical validity). The simplification is
only that she does not know of any step that would be justificatory by her
own lights for the conclusion a, and that makes do with some proper subset of
{b , . . . , bn } as its premise set. Suppose, however, that, unknown to the agent,
the subset {b , . . . , bn } furnishes adequate logical support for athat there is a
logical proof of the sequent {b , . . . , bn }, therefore a. Then, as it happens, she
might contract with respect to a in a strategically disastrous way, by surrendering b and thinking (mistakenly) that she had thereby done enough to ensure
that a could not be justified by appeal to such beliefs as remain in her belief
scheme. In the subjective, agent-relative sense, she is right; for, ex hypothesi, she
does not know of any justificatory transition from {b , . . . , bn } to a. Her continuing belief in b , . . . , bn , as it happens, logically commits her (in the appropriately
objective, agent-independent sense) to belief in abut she does not know that.
In such a case we cannot hold the agent accountable to the external, objective,
transmission of warrant, for she does not know of it. From within her own
perspectiveby her own lights, as we have been sayingshe has done enough,
in giving up b , to neutralize the step {b , . . . , bn }, therefore a in the role it
had played in her earlier belief scheme. While the belief a could be justified (by
external standards) by appeal to such beliefs as remain in her belief scheme, it
cannot be justified by the appropriately internal standards, which involve appeal
only to such steps as are known to the agent.
Rational agentslogical paragonsare constantly hostage to the historical fortune of discovery of proofnot its mere Platonic existence. All we can
demand of paragons is that they carry out their contractions and revisions in a
rational fashion in accordance with what they already know of logical relations
as a result of their past discovery of proofs.
Given any set of premises {b , . . . , bn }, and conclusion a, there can be only
one inference stroke mediating the step from {b , . . . , bn } to a. (For the transition is a matter of those premises supporting that conclusion.) This means that,
in the formal modeling presented in Chapter , the step can be represented as
the set
{{b , . . . , bn }, a}.
This extensionality condition on steps has its diagrammatic counterpart: any
two distinct inference strokes receiving arrows from exactly the same nodes
must send their single outgoing arrows to respectively distinct nodes. Remember, every inference stroke sends out exactly one arrow (that is, all steps are
NODES AND STEPS
45
A X I O M S F O R A B E L I E F N E T WO R K
47
Nb : Nb
Sb : Sb
Nb , Sb : Nb Sb
Acb : Acb Nb , Sb : x(NxSx)
Sc : Sc
axiomatization at first order, the strokes are also treated as individuals in the
domain. But not even the two-sorted axiomatization at first order treats the
arrows of the diagram as individuals. Rather, the arrows collectively represent
the extension of the binary relation A among nodes and strokes.
49
4. Every white inference stroke that receives an arrow receives an arrow from some
white node.
x((Wx Sx zAzx) y(Wy Ny Ayx))
(x is white) with B(x). If we take this course, then Axioms () and () would
become, respectively,
1B . Everything is either black or not black
x(Bx Bx)
0B . Nothing is both black and not black
x(Bx Bx)
So on this approach (B ) comes out as an instance of the Law of NonContradiction, and (B ) comes out as an instance of the (classical) Law of
Excluded Middle for monadic predicates. The latter does not offend against
intuitionistic scruples, since we readily concede that the predicate B(x) is decidable. It is for this reason that we are not unduly concerned to make special
mention of Axioms () and (). The closer analysis just given renders them
superfluous. We concentrate instead on the other four Axioms of Coloration,
which is where the action is.
51
Our main Axioms of Coloration ()() rule out the following arrangements.
V I O L AT I O N O F A X I O M ( 1 )
If the only inference strokes sending arrows to a were those shown, then a
should be white, not black, on pain of violating Axiom (): Every black node
52
receives an arrow from some black inference stroke. The agent is represented as
believing a, but with no active justification or justificatory path leading to a
(that is, with no arrow coming to a from a black inference stroke, even if only
an initial one).
V I O L AT I O N O F A X I O M ( 2 )
Here we have a violation of Axiom (): Every white node receives arrows (if any)
only from white inference strokes. It is a failure of closure under known steps.
The agent is represented as apprised of the step from the two premises to the
conclusion, as believing the premises (black) but as not believing the conclusion
(white).
V I O L AT I O N O F A X I O M ( 3 )
A simple change of blacks and whites produces our next configuration to illustrate axiom violations:
A X I O M S F O R A B E L I E F N E T WO R K
53
Here we have a violation of Axiom (): Every black inference stroke receives
arrows (if any) only from black nodes. The step is being taken (with black inference stroke and black conclusion) as providing justification for the conclusion;
while yet one of the premises (white) is not believed.
V I O L AT I O N O F A X I O M ( 4 )
Another change of blacks and whites produces our final axiom-violating configuration:
Here we have a violation of Axiom (): Every white inference stroke that receives
an arrow receives an arrow from some white node. It is a failure of closure
(downwards) under a known step. The premises are believed, and the step is
known. So the inference stroke should be black, and the conclusion should
be too (since, by the agents own lights, it is to be believed). Alternatively: it
is a failure of closure (upwards) under a known step. The conclusion is not
believed, and the step is known. So at least one of the premises should not be
believed.
Which of these two alternatives prevails will depend on whether one is busy
expanding or contracting the belief network. If expanding, the former alternative prevails; if contracting, the latter one prevails. This is not merely a case of
one agents modus ponens being another agents modus tollens. Rather, the fix
for the violation will depend on the agents purposes at the time.
Similar remarks apply to the violation of Axiom () above. The fix in each of
these cases will depend on whether one is intent on spreading black (expanding) or on spreading white (contracting).
A X I O M V I O L AT I O N S W I T H I N A N A R ROW E R
D I AG R A M M AT I C A L C O N T E X T
as its topmost elements and only nodes as its bottommost elements. Here, we
shall repeat the exercise of picturing the four possible violations, but focus on
a narrower diagrammatical context in doing so. We shall display only those
fragments of diagrams that exhibit the violations. In doing so, we shall allow
inference strokes to appear either as topmost or as bottommost elements of
diagrams.
Here now are the same axiom-of-coloration violations as before, but pictured
more economically.
V I O L AT I O N O F A X I O M ( 1 )
all white
...
Here we have a violation of Axiom (): Every black node receives an arrow from
some black inference stroke.
V I O L AT I O N O F A X I O M ( 2 )
Here we have a violation of Axiom (): Every white node receives arrows (if any)
only from white inference strokes.
V I O L AT I O N O F A X I O M ( 3 )
...
...
Here we have a violation of Axiom (): Every black inference stroke receives
arrows (if any) only from black nodes.
A X I O M S F O R A B E L I E F N E T WO R K
55
V I O L AT I O N O F A X I O M ( 4 )
al l black
...
Here we have a violation of Axiom (): Every white inference stroke that receives
an arrow receives an arrow from some white node.
Diagram A
56
The white steps (identified by having white inference strokes) are known to the
agent, but cannot contribute to the closure of her belief system (the black nodes).
This is because with a white step, the agent does not believe all its premises: at
least one premise is white. All it takes is one non-believed premise to render
a known step irrelevant for purposes of closure. Thus, even if a white step
has some black premises, its conclusion will still be white (unless the node in
question stands also as the conclusion of some other step, all of whose premises
are black).
With belief systems represented this way, the problem of contraction is particularly vivid. Suppose someone points to a black node, and orders that it be
changed to white. In other words, the agent is being told to suspend that belief.
All the transitional steps are to remain in place; only the colors of nodes and
inference steps are to change (as a result of compliance with this order). How
should such changes be effected? The agent must seek to minimize the spread
of white, and salvage a maximal amount of black.
The result of spreading white must conform to Axioms ()(). Moreover, the
process will be subject to the following overarching constraint:
[White Lock] Do not make anything white black.
The constraint is in place because we are concerned, with contraction, only with
cases where the agent has to give up certain beliefs, and not take on any new
ones.
The basic actions that can or must be taken in this process are of the following
Action Types:
Action Type . A black node that is initial (i.e. believed outright) can be made
white, and made to stand unsupported by any (black) inference stroke above it:
If a were an observation report (this is a brown penny, for example) one could
imagine the agent ceasing to believe a upon discovering, say, that the lighting
conditions had been abnormal.
S P R E A D I N G W H I T E V. S P R E A D I N G B L AC K
57
This way of representing the change conveniently assimilates the initial case
to the general case where the node a might be justified by at least one set of other
nodes (possibly even: a plurality of such sets). See Section .. for further discussion of this possibility. As already remarked in Section .., calling a node
initial does not mean that it cannot also stand as the conclusion of a justifying
step involving other nodes. In Section .. we shall say more about this.
Note that the reverse of Action Type is not an option during a contraction
process. Contraction has only to do with surrendering beliefs, not adopting
them. The constraint White Lock on contractions makes this explicit.
Action Type . This type of action modifies a diagram so as to move from a situation in which certain epistemic norms are violated to a situation that conforms
with those norms. Here, the norms in question are our Axioms of Coloration,
which ensure that a dependency network can be interpreted as the belief scheme
of a rational agent. A black step whose (formerly black) conclusion (Subdiagram
a) has been made white (Subdiagram b) must have its inference stroke, and
at least one of its premises, made red (Subdiagram c). Subdiagram b is normviolating; but Subdiagram c conforms with the norms.
(this one, say)
choose a premise
for whitening
Subdiagram 2a
Subdiagram 2b
Subdiagram 2c
One has to disable the premise set, in order to prevent the conclusion from
being forced to be black by the step in question. This is called an upward
sweep (of whitening). It is roughly like performing a modus tollens on a known
step. We say roughly like because what is being transmitted back (or up) from
conclusion to premises is not falsity or disbelief, but rather lack of belief. That is
the status represented by a nodes being white.
The only problem is: Which premise do we choose for whitening? Philosophers
call this the QuineDuhem problem.2 Computer scientists will recognize it as
a source of non-determinism of any algorithm for belief-scheme contraction
58
within this formal modeling that we are providing. Note that Subdiagram b
represents the first stage of a (local) contraction process, and does not comply
with Coloration Axioms ()(). Indeed, it violates Axiom (): Every white node
receives arrows (if any) only from white inference strokes. The further whitening
in Subdiagram c rectifies this situation. It does so by whitening the inference
stroke, and the right-hand node shown. Of course, the left-hand node could
have been whitened instead of the right-hand one, to the same effect (compliance with Axioms ()()):
(this one, say)
choose a premise
for whitening
Subdiagram 2a
Subdiagram 2c
Subdiagram 2b
S P R E A D I N G W H I T E D OW N WA R D S
Action Type . A black step (Subdiagram a) any one of whose (formerly black)
premises has been made white (Subdiagram b) must have its inference stroke
made white (Subdiagram c):
Subdiagram 3a
Subdiagram 3b
Subdiagram 3c
Note that Subdiagram b represents the first stage of a (local) contraction process, and does not comply with Axioms ()(). Indeed, it violates Axiom ():
Every black inference stroke receives arrows (if any) only from black nodes. The
whitening of the inference stroke in Subdiagram c rectifies this situation. For, if
one of the premises of a transitional step is white (i.e. not believed), then those
premises do not, collectively, suffice to justify belief in the conclusion. (This
is not to say, however, that there can be no other sufficient justification for the
conclusion in question. It is only to say that this particular step, whose premises
are not all black, does not by itself justify belief in the conclusion.)
S P R E A D I N G W H I T E V. S P R E A D I N G B L AC K
59
Action Type . As we have just observed, the process in () above does not
necessarily entail that the conclusion of the step (call it a) must also be made
white. That is required only when every step (known to the agent) of which a
is the conclusion has a white inference stroke, as in the belief-scheme fragment
illustrated below. Here, we show three inference strokes altogether that send
arrows to node a. These represent three different ways in which the agent could
take herself to be justified, or would be justified, in believing aif only one of
those strokes were black.3 But they are all whitein particular, the one in the
middle. For the strokes on the left and on the right, we omit the premise nodes
that send arrows to them; for the reader is being asked to focus on the step in
the middle, the changing coloring of whose premise and conclusion nodes is
what is salient for our present purposes:
a
Subdiagram 4a
a
Subdiagram 4b
This is called a downward sweep (of whitening). If the belief a has been
deprived of all the justificatory support that it formerly enjoyedthat is, when
every inference stroke sending an arrow to a has been made whitethen a must
be given up (must be made white). Subdiagram a does not comply with our
Axioms ()(); for it violates Axiom (): Every black node receives an arrow
from some black inference stroke. The whitening of node a in Subdiagram b
rectifies this situation.
W H AT C O U N T S F O R C O N T R AC T I O N
When one is contracting a belief system (turning black inference strokes and
nodes to white ones), it turns out that only the steps that are black at the very
beginning of the contraction process are relevant. Put another way, one could,
With this contrast between an agent taking herself to be justified versus actually being justified,
we are indicating that our dependency networks could admit of either an external, objective
reading (regarding matters of epistemic justification) or an internal, subjective reading. In this
exposition, we systematically prefer the latter reading. But it is worth keeping in mind the possibility
that the formalism of dependency networks could be put at the service of a more objectivist
epistemology also. This point was made by Salvatore Florio.
60
before the contraction process even begins, expunge from the diagram every
white node and white inference stroke, and every arrow leading to or from a
white inference stroke. The absence of these elements will not affect what the
outcome of the contraction process will be. This is because contractions go to
work only on the black part of a diagram, turning some of it to white. And the
considerations that prompt the spread of white within the erstwhile black are
not sensitive to any of the white within the original diagram. Thus, if one were
contracting the (black) belief scheme within Diagram A:
Diagram A
one could begin by limiting ones attention just to the black steps therein:
Diagram A Black
S P R E A D I N G W H I T E V. S P R E A D I N G B L AC K
61
We have labeled three of the nodes a, b and c for the following purposes. Suppose, for the sake of illustration, one were ordered to contract this belief system
with respect to the node a. If one accordingly gives up b, then one obtains the
contraction on the left; whereas if one decides instead to give up c, then one
obtains the contraction on the right:
As the reader can easily verify, each of these contractions is in conformity with
Axioms ()(). Moreover, each counts as a minimally mutilating contraction,
by virtue of being maximally black without blackening the node a, in the following sense:
If one were to extend the blackening in any way within the diagram, then, in making the
result conform with Axioms (1)(4) (under the constraint Black Lock, to be explained
presently), one would be forced to make the node a black again.
The two diagrams above represent the outcomes of two possible spreads of
white emanating from the whitening of a, in pursuit of conformity with Axioms
()() (under the constraint White Lock). All that was relevant for this process
was the original extent of black. The two divergent outcomes reflect the multiplicity of choices that could be made along the way. In the contraction process,
we had some of the original black turning to white, with White Lock in effect.
The nodes and inference strokes that were originally white were irrelevant for
this further whitening. They could not have affected our choices in any way.
The original white nodes and inference strokes do become relevant, however,
as soon as one expands a belief scheme (or any of its contractions) by adopting
new beliefs in such a way as to make any formerly white node become black.
62
(One would also, thereby, put Black Lock into effect.) This process could well
turn a formerly white inference stroke into a black one, upon all the premises
of the step in question becoming black. In that case, the conclusion of the step
would also become black, and so on.
Consider now the belief scheme in Diagram (i) below, which conforms to
Axioms ()().
a
Diagram (i)
The belief system is to be contracted, say, with respect to its element a. This is
indicated in Diagram (ii), in which a is now whitened. Also, remember that with
contraction the constraint White Lock is in effect.
a
Diagram (ii)
Diagram (ii) violates Axiom (): Every white node receives arrows (if any) only
from white inference strokes. We have to rectify this situation. Two Action Types
are available to do so. We could make a black again; or we could whiten the
inference stroke above a. The first alternative would be, in effect, to abandon the
process of contraction with respect to a supposedly embarked upon. (It would
S P R E A D I N G W H I T E V. S P R E A D I N G B L AC K
63
a
Diagram (iii)
Alas, Diagram (iii) violates Axiom (): Every white inference stroke that receives
an arrow receives an arrow from some white node. Again, two alternative Action
Types present themselves: we could make the inference stroke black again,
thereby returning (ill-advisedly) to Diagram (ii), in violation of White Lock; or
we could bite the bullet and whiten one of the premises supporting a. Suppose
we choose to whiten the leftmost premise, b. The result is Diagram (iv.a), which
at last conforms to Axioms ()().
a
Diagram (iv.a)
The initial whitening of a has induced the upward whitening of the inference
stroke above it, as well as the whitening (for Diagram (iv.a)) of the leftmost
supporting premise b. The only black bit of the original diagram that survives
in this contraction is the outright belief in d.
Suppose now that we choose (in Diagram (iii)) to whiten the rightmost
supporting premise d (for the step with conclusion a) rather than the leftmost
supporting premise b. Then we would obtain Diagram (iv.b):
64
a
Diagram (iv.b)
We point out the alternatives (iv.a) and (iv.b) in order to underscore once
again the phenomenon of non-deterministic choice with our algorithmic
method of contraction. Whenever a conclusion a of a step is being whitened (i.e.
surrendered), and the step has more than one premise, there are correspondingly many ways of inducing upward whitening (from a) at the next stage. One
needs to disable the premise set in question; but, in order to do so, it suffices to
whiten only one of its premises. Moreover, that remark applies to each step (in
the dependency network being contracted) that has a as its conclusion. Hence,
if any of these premise sets overlap, one might consider whitening a common
premise, thereby minimizing the amount of whitening induced. This aspect of
the procedure will be discussed in greater detail in Chapter .
In order to illustrate this constraint in action, let us return to the stage reached
above in Diagram (iv.a):
a
Diagram (iv.a)
S P R E A D I N G W H I T E V. S P R E A D I N G B L AC K
65
Imagine now that the agent acquires outright belief in the sentence labeled c,
which thus far has been white. We make both the node c and the inference stroke
above c black, to obtain Diagram (v).
a
Diagram (v)
We are not yet done. Diagram (v) violates Axiom (): Every white inference
stroke that receives an arrow receives an arrow from some white node. The rightmost two-premise step is the new culprit: it has black premises (d and c), but
a white inference stroke and a white conclusion. Once again, two alternative
Action Types are open to us. We can whiten one of the premises; or we can
make the inference stroke (and thereafter the conclusion) black. Whitening
either one of the premises is counterproductive. It would violate the constraint Black Lock on the process of expansion. For, if we whiten c, we shall
be abandoning the very expansion begun by the blackening of c. Moreover, if
we whiten the other premise d, we shall not be conserving as much as possible of the original black in Diagram (i). So, we decide (in conformity with
Black Lock) to make the inference stroke, and the conclusion below it, black.
The result is Diagram (vi) below, which conforms to Axioms ()(); so we
are done.
a
Diagram (vi)
66
belief scheme
The foregoing example shows that the process of closure (with respect to known
steps), upon which it is rational to insist, can make formerly white steps (i.e.
the ones with white inference strokes) significant for an expansion of a contraction. That is why, in a general theory of revision of belief schemes, which
involves contraction followed by expansion, it is important to make provision
for both the white and the black parts of those transitional steps known to the
agent. For the revision of a (black) belief scheme with respect to the (formerly
white) proposition p is effected by first contracting the scheme with respect to
p (here we are assuming that the scheme contains p, whose node would
of course be black at the outset); secondly, expanding the contracted result
by adjoining p; and thirdly, closing with respect to the steps known to the
agent.
In borrowed notation familiar to students of the belief-revision literature, we
take a scheme B containing p. We contract B with respect to p, so as to obtain
the contraction
(B p).
Then we adjoin p, and take the closure of the result. This operation is called
expanding with respect to p. In the case at hand, we obtain the expansion
[(B p), p].
(Some authors write this expansion as (B p) + p.) Note, however, that there
are significant differences between what these symbols denote in AGM-theory,
and what they denote according to the present account. For the AGM-theorist,
B is a theory in a formal languagean infinite set of sentences closed under
logical consequence (or deducibility). On the present account, B is a finitary
belief scheme, consisting of steps made up of nodes and inference strokes,
whose white or black coloration reveals what is believed by the agent, and
how, and why (as far as she is concerned). For the AGM-theorist, the contraction function need not be computable; on the present account, it must be.
(This is a self-imposed constraint, for the purposes of cognitive modeling.4 )
T H E R O L E S O F T H E W H I T E A N D B L AC K PA RT S
67
For the AGM-theorist, taking the closure [(B p), p] means taking the (not
necessarily decidable, or even effectively enumerable) logical closure of the set
(B p) {p}, which in general is a non-effective operation; whereas on the
present account it means effectively determining a new finitary belief scheme
from an old one.
One can now appreciate why it is important to conserve all (and only)
the (most economical) transitional commitments that the agent has already
acquired, even if much of it might lie within the white part of the overall scheme
before any revisions take place.
68
Contraction
under White Lock:
Shape of Fork
Expansion
underBlack Lock:
... ...
Stable
Stable
Stable
Stable
Stable
Stable
Stable
Stable
...
...
... ...
...
...
Does not arise
... ...
... ...
Does not arise
... ...
...
...
... ...
... ...
...
69
2.4.2 Example 1
This example is a simplification of one to be found in Hansson [], at
pp. . It is also an instance of the schematic example given by Fuhrmann
[], at p. .
The agent believes that
(a) Cleopatra had a daughter.
So the agent believes b on the basis of a (Subdiagram a). But the agent subsequently ceases to have any confidence in the source of the would-be information
that Cleopatra had a daughter. The agent therefore gives up the belief a. (Subdiagram b below: the node a is now white, and without any support.) So, because
a was the sole support in the agents belief scheme for b, the agent gives up the
belief b as well (Subdiagram c). Now, if the agent were somehow to come across
reliable but non-specific information to the effect that Cleopatra did, after all,
have a child, but with the sex of the child left unspecified, the node b could
become black again, this time with its own initial (black) inference stroke, and
without re-instatement of the belief a (Subdiagram d):
Subdiagram 1a
Subdiagram 1b
b
Subdiagram 1c
b
Subdiagram 1d
2.4.3 Example 2
This example comes from Hansson [], at pp. , concerning a belief
set K. (The reader should be advised that we use brackets to indicate either
useful interpolations, or innocuous changes of notation that are intended to
ensure consistency of exposition. So the brackets within this quote do not
indicate logical closure!)
I previously entertained [at Stage 1] the two beliefs George is a criminal () and George
is a mass murderer (). When I received information [i] that induced me to give up the first
of these beliefs (), the second () had to go as well (since would otherwise follow from
) [I thereby reached Stage 2]. I then received new information [j] that made me accept the
belief George is a shoplifter (). The resulting new belief set is the expansion of [K ]
by , [(K ) + ]. [Call this Stage 3.] Since follows from , [(K ) + ] is a subset of
[(K ) + ]. By recovery, [(K ) + ] includes , from which follows that [(K ) +
] includes . Thus, since I previously believed George to be a mass murderer, I cannot
any longer believe him to be a shoplifter without believing him to be a mass murderer.
Clearly the pieces of information i and j (whatever they might have been)
conflict with one another, as does the piece of information i with the claim .
Let us assume that from the very beginning the agent realizes that being a
shoplifter entails being a criminal, even though (at the outset) the agent does
not believe that George is a shoplifter. (A mass murderer, yes, but at any rate not
also a shoplifter!) Here, then, is the sequence of changes described in Hanssons
counterexample:
Stage 1
Stage 2
(on learning i)
Stage 3
(on learning j)
T H E R O L E S O F T H E W H I T E A N D B L AC K PA RT S
71
Each of these three stages complies with Axioms ()(); we do not show the
intermediate stages of the two revision processes from Stage to Stage and
from Stage to Stage .
2.4.4 Example 3
This example comes from Niedere []. Once again, let K be a belief set,
and suppose that K implies A. Suppose also that the agent is apprised of the two
transitional steps A, therefore A B and B, therefore A B.
AB
AB
AB
..
.
b
2.4.5 Example 4
This example comes from Levi [], pp. . We quote at length, with
B A in place of Levis distracting contrapositive A B:
Suppose that we know that Jones suffers from hepatitis and that he contracted it by eating
tainted humus [sic].5 Call Jones ate tainted humus A and Jones suffers from hepatitis B.
Consider the contraction of our corpus K by giving up B. We would be faced with giving
up either A B or A B. We might quite plausibly regard giving up the claim A B
as incurring greater loss of informational value than giving up A B on the grounds
that in doing so, we would be giving up the assumption that eating tainted humus tends
to result in becoming infected with hepatitis, which has some explanatory value.[fn] But
then if we retain A B, we shall be required to give up A. The only question that remains
is whether, in giving up A, we should also give up [B A] or retain it. But if we put
ourselves in the position of someone who does not know that Jones has hepatitis and
does not know that he has eaten tainted humus, would we take ourselves as knowing that
if the person has hepatitis, he has eaten tainted humus? There are contexts in which we
would respond affirmatively. We would do so if we had information that ruled out that
Jones had sexual contact with an infected person, used tainted needles, undergone blood
transfusions, and the like. Suppose, however, that we lacked such information. If the initial
corpus that contained the information that Jones had eaten tainted humus did not rule
out that he had sexual contact with someone infected with hepatitis or some other such
source, then when B was removed and subsequently re-instated, we would not be in a
position to conclude that Jones had eaten tainted humus. But this means that [B A]
T H E R O L E S O F T H E W H I T E A N D B L AC K PA RT S
73
is not retained in the contraction. It seems to me that in situations such as the one
envisaged, we do not think that the loss of [B A] adds to the loss of informational value
incurred in giving up B additional to what is lost by giving up A B even though it adds
to the loss of information. The extra information lost is considered irrelevant. We might
say that we initially believed [B A] only because of our initial conviction in the truth
of A. No loss of explanatory information is incurred by giving up [B A] except that
incurred in giving up A.
AB
BA
AB
Levis point is that contraction of this belief scheme with respect to B should
result in ones giving up BA as well.
The point of interest here, though, over and above our agreement with Levi on
this score, is how the contraction would proceed with respect to B, and what the
final coloring of the contracted scheme would be. First, as a result of whitening
B, and the desire to salvage belief in the explanatorily prized generalization C,
we would decide to whiten A (since, if we whitened AB, we would have to
whiten C). Secondly, since both A and B are now whitened, we would whiten
the two conditionals below them, which have no other support in the diagram:
A B
B A
A B
T H E R O L E S O F T H E W H I T E A N D B L AC K PA RT S
75
2.4.6 Example 5
Our next example is the one given in Tennant [a], at p. . It involves
the two outright beliefs A and AB, along with their two consequences BA
and B:
<
BA
AB
As the reader will easily detect, this is a subdiagram of the diagram above
depicting Levis example. At the time of writing the BJPS paper the author
was not, alas, aware of Levis reflections on tainted hummus, unsafe sex,
shared needles, blood transfusions and hepatitis. But even after encountering
his example, it was not until we analyzed it in the diagrammatic detail just
given that we realized that it properly contained the subdiagram just isolated
above. Our own example was couched directly in terms of the simple formulae
and steps just depicted, and with the permissible assumption that for the agent
concerned, the conditional AB was more entrenched than its antecedent A
(as indicated in the diagrams immediately above and below by the sign <).
Suppose one is to contract with respect to B. Then A will have to go. So too
will BA, since it is believed only because A is believed. Upon restoring B, we
will not get A back again:
76
Result of contracting
with respect to B:
<
BA
AB
<
BA
AB
2.4.7 Example 6
Our final example comes from Quine and Ullian [] at p. . The following
dependency network illustrates the forensic predicament they describe. The
authors do not tell their reader where the murder was committed; so let us
assume that it was in Boston.
nfe
nfb
abc
77
A
B
C
nfe
nfb
a
b
c
The network is drawn so that doubt can fall only on its nodes, not on its
transitional steps (marked with the horizontal strokes). The network is unacceptable in the current state, because its final step has conclusion , from four
premises indicated as believed (colored in black). Our task is to whiten parts
of the network, so as to restore consistency. Whitened nodes will no longer be
believed. The problem is: what is the most economical and sensible way of doing
this (i.e. spreading white upwards from the bottommost node, labeled )?
As Quine and Ullian tell the story, ones response to the terminal contradiction should be to reject [b] or [a b c], thus either incriminating Babbitt
or widening our net for some new suspect. We shall illustrate here the effect of
following either of these courses of action.
First, suppose we reject b. Then the network would become
78
nfe
nfb
abc
The reason why this would now incriminate Babbitt is that one credits
the agent with the ability to make the easy propositional inference from the
premises
a, c, a b c
to the conclusion b. The step that would represent this inference is not shown
explicitly in the diagram. It is interesting to note that the conclusion b is not
vouchsafed (as black) simply on the basis of the whitening of b. It is only
because the step just mentioned:
a
abc
nfe
nfb
abc
T H E R O L E S O F T H E W H I T E A N D B L AC K PA RT S
79
These two ways of contracting are dictated by the fact that the beliefs B
and nfb (no fourth beneficiary) are the least entrenched among the initial
nodes. Without recourse to such entrenchment considerations, it would be just
as licit to contract in a way that casts doubt on the reputation of the Albany hotel:
nfe
nfb
abc
80
is of some interest, for two reasons, which have already been broached in
Section ..
First, although NP-completeness is an indication to computer scientists that
a problem is intractable, it is nevertheless the lowest degree of computational
complexity to which anyone can aspire when doing anything reasonably serious
in computational logic. Even the decision problem for truth-table satisfiability
in classical propositional logic is NP-complete. (Decision problems for theoremhood in other propositional systems are much worsesee Section ..)
Thus we can claim that our contraction method is no more complex than a
truth-table test.
Secondly, since one has to consider sets of sets of premises in the contraction
process, there was the prima facie danger that the complexity of the problem
might exponentiate above that of a truth-table test. Fortunately, however, this
turns out not to be so.
A computational treatment was made possible only by conceiving of the
inputs (the belief systems) as finitary objects, and by isolating simple, mechanically implementable rules for their manipulation so as to produce outputs
with a certain decidable property. It is understandable that one might therefore be concerned at the very outset that such a simple conception of the
matter must do some violence (albeit in the interests of effectiveness) to the
various underlying logical notions. After all, are there not theories that are
not finitely axiomatizable? What about undecidable theories? And is not the
logical closure of even a finite axiomatic basis itself infinite? Are not these
logical infinities and undecidabilities going to pose insuperable problems to an
effective (hence finitary) method of theory contraction that aspires to be fully
general?
These concerns will be addressed and allayed in Chapter . Even an infinitely
axiomatized and undecidable theory will only ever have been finitely developed
by the time any contraction might be contemplated. (This notion of finite development will be discussed in greater detail in Chapter .) Deductions, or steps,
that a rational agent does not yet know about are perforce irrelevant to what
she ought to do upon giving up any statement that she knows to be one of her
present commitments. The diagrams above isomorphically depict these finite
dependency networks, and contain all the information relevant to a rational
process of theory contraction. Moreover, they do so in a way that makes the
whole process algorithmic, and of the lowest complexity that could reasonably
have been expected.
S U M M A RY A N D C O N C L U S I O N
81
82
CHAPTER 3
Global Conditions on
Contraction
theories
Ramanujan, the great mathematician, claimed that an angel visited him in his
sleep at nights, and wrote upon a tablet various theorems that Ramanujan ought
to prove.1 Ramanujan would dutifully prove them; and his contribution to the
expansion of our knowledge in mathematical analysis was considerable.
The story about the goddess of Namakkal is familiar from C. P. Snows foreword to the
reprinting of Hardy []. But much richer detail is available in Robert Kanigels more recent and
extraordinarily thoroughly researched biography of Ramanujan. From Kanigel [], at p. :
In Ramanujans family, the family deity was the goddess Namagiri, consort of the lion-god Narasimha.
Her shrine at Namakkal was about a hundred miles from Kumbakonam [the town where Ramanujan
grew upNT ], about three-quarters of the way to Erode, near where Komalatammals [Ramanujans
mothersNT ] family came from. ...Ramanujan ...would utter Namagiris name all his life, invoke
her blessings, seek her counsel. It was goddess Namagiri, he would tell friends, to whom he owed
his mathematical gifts. Namagiri would write the equations on his tongue. Namagiri would bestow
mathematical insights in his dreams.
But the available evidence does not exactly identify a unique supernatural agency at work. We learn
also, ibid. at p. :
T. K. Rajagopolan, a former accountant general of Madras, would tell of Ramanujans insistence that
after seeing in dreams the drops of blood that, according to tradition, heralded the presence of the
K N OW L E D G E S E T S A N D B E L I E F S E T S V E R S U S T H E O R I E S
83
84
G L O B A L C O N D I T I O N S O N C O N T R AC T I O N
more clearly why the result holds, or that lend themselves to generalization in
the solution of whole classes of related conjectures. The upshot, then, is that a
typical item in a knowledge (or belief) scheme is a known (or believed) proposition furnished with possibly more than onebut still only finitely many
justifications, or epistemic pedigrees, or proofs.
One can have different bases for one and the same theory; indeed, one can
have different irredundant bases for one and the same theory. For take the two
irredundant bases {a, b} and {a, a b}. From the first of these bases one can
deduce a b by means of one step of I; and from the second, one can deduce
b by means of one step of E. In each case that yields the same development
BASES, DEVELOPMENTS AND THEORIES
85
{a, b, a b}; and from there on everything will coincide, and the same unique
theory will be generated.
This example serves also to illustrate what we shall mean by a development.
In the case where the base is finite, a development will be any (finite) set lying
somewhere between that base and its full logical closure. To be realistic, a
development moreover is always a finite set:
(finite) base
{a, b}
(finite) development
{a, b, a b}
G L O B A L C O N D I T I O N S O N C O N T R AC T I O N
3.3 Contractions
Back to Ramanujannot the historical one, but rather a fictional extension
of him for expository purposes, whom we shall call Ramanujan . Suppose
Ramanujan is interested in theories and belief sets quite generally, and not just
mathematical ones. (What we shall be saying, however, will apply to mathematical theories as well as to non-mathematical ones.) Ramanujan s angel now
comes to him in his sleep holding a different kind of tablet. No longer does
the angel write down whatever mathematical claim p he wants Ramanujan
to believe, and instruct him to find a reason for believing p (to find, that is,
a proof of p). Rather, the angel writes down what he wants Ramanujan to
cease believing. Indeed, when the angels practice changed, he had an initial
conversation with the dreaming Ramanujan about the new tack. The angel
spake thus:
In the past I have told you what propositions are worth believing. But from now on I
shall tell you what propositions you should no longer believe. Do not, however, leap to any
conclusions about their contradictories. Im not telling you to disbelieve them; I just want
you not to believe any proposition I may happen to write on the tablet. But, as far as each
such proposition will be concerned, I want you to keep an open mind. For who knows?
one day you might have to recruit back the beliefs Ive asked you to let go. Or, you may
have to take on some of their contradictories. But, for the time being, I cant say which. So
do please keep an open mind. Obey my injunctions by giving up as little as possible. And
if ever I instruct you to cease believing a logical truth, you can disregard that instruction.
Take seriously only my proscriptions of propositions that are not logical truths!
Ramanujan has been assigned the general task of contracting a set of beliefs
with respect to any given (contingent) proposition implied by it. A contingent
proposition is one that is neither logically true nor logically false. Contracting
with respect to a logical truth involves doing nothing. And the reason why
contraction with respect to a logical falsehood is not explicitly considered is
that one assumes an abiding commitment to consistency of belief anyway, in
the form, perhaps, of the angelic proscription Do not believe !
Ramanujan falls to reflecting on how best to carry out the angels commands.
He knows the shape of the problem in general: at any stage he will have a set A of
current beliefs. There will be only finitely many of them; but, potentially, at least,
there is an infinite set of beliefs derivable from A. This is because the believer is
rationally committed to any proposition that can be shown to be in the logical
closure of A. (Among these, for example, will be all the logical truths in the
C O N T R AC T I O N S
87
88
G L O B A L C O N D I T I O N S O N C O N T R AC T I O N
over a; whence one will give up a. Thus one gives up both a and b regardless of which of
these two one gives up first. The residue will therefore be [a b].
89
(or lemma) for the ultimate proof, from first principles, of the corresponding
result a.
Note that the representation of above could also be furnished as the following list of steps with conclusion a:
, a, . . . , n , a.
And indeed this will be our preferred method of representation, within the
modeling to be described in Chapter . For one finds that the most convenient
kind of object on which to focus, for formal manipulations, is the step. It is better
to treat the various steps just displayed above severally and individually, as it
were, rather than to treat the set
{ , a, . . . , n , a}
of steps that have a given node a as their common conclusion.
It is an interesting question, and one worth exploring, whether making do
with the most recent lemmata (as members of i ) adduced for the proof(s) of
a, rather than tracing logical dependencies all the way back to those members
of the base in question on which a ultimately rests, substantially affects the
outcome(s) of the contraction process. If it did, then this would be yet another
manifestation of the intensionality of that process. This question is one that is
unable to be posed on the AGM-approach, since it does not model the dependencies or pedigrees of justification to which we have been drawing attention.
Ramanujan knows that the angel could write on his tablet any proposition p,
regardless of whether it features as an already justified item in the currently
developed scheme A. If there is as yet no justification, within A, of the prohibited
proposition p, then Ramanujan will have to live in constant fear of the prospect
that his scheme might unexpectedly turn up such a would-be justification for p,
contrary to the angels wishes. Ramanujan will just have to be very vigilant to
avert such a prospect; but, apart from wondering about what proofs of p there
might, unknown to him (at present) be, there is precious little that he can do,
for the time being, to respect the angels wish that he not believe p. If there is
already a proof for p in his scheme A, however, so that A has a justified item
of the form
= p, {. . . , , . . .}
then matters are different. Ramanujan in this case has his work cut out for
him; for he has to excise p and disable every one of those justifications like .
Fortunately, the cruder form of representation
90
G L O B A L C O N D I T I O N S O N C O N T R AC T I O N
= p, {. . . , , . . .}
(where is the set of premises of the justification ) suffices for this doxastic
surgery. For all that Ramanujan has to do is remove some member, or members, from each such set , so as to make what is left over insufficient for the
justification of the proscribed proposition p.
Exactly how Ramanujan will contrive to do this is a matter we can leave,
at this stage of our tale, to his own sharp wits. We have already begun the task
of presenting a detailed method, which we have just illustrated above with the
toy examples of contracting [a, b] and [a, a b] with respect to a b. Let us
remark, however, on one stratagem that Ramanujan considers and quickly
rejects.
W H AT D O W E R E Q U I R E O F A C O N T R AC T I O N ?
91
p itself? Surely an open mind means that one can proceed either to adopt p
or to adopt p as a new belief without thereby incurring any inconsistency?
But if Ramanujan were to pursue the cheeky intuitionist path, he would be
unable subsequently to embrace p without thereby engendering conflict with
his belief p, even in the absence of p.
Realizing this, Ramanujan sees that one has to be careful when formulating
the conditions for success in the project of contracting a set A of beliefs with
respect to any of its consequences p. The contraction A p (that is, the kernel
of the contracted scheme ) must not only fail to imply p; it must also be consistent
with p (that is: A p, p ). Against the background of classical logic,
these two requirements are equivalent; against the background of intuitionistic
logic, however, they are not. In the intuitionistic case the way to express the
condition for success is to insist that the contraction A p be consistent with
p. This will guarantee that A p fails to imply p. Intuitionistically, merely
insisting that A p fail to imply p is not enough to guarantee that A p will
also be consistent with p.
Let us now be mindful once again of the distinction drawn earlier between
knowledge sets and belief sets, on the one hand, and knowledge schemes and
belief schemes on the other. A belief set is just the set of propositions believed
by an agent. But her belief scheme provides for each belief in that set a list of
justificatory pedigrees within her system of belief. This is the extra dimension
that underlies some of the intensional behavior of beliefs in a contraction
process.
The task of contracting a belief scheme A with respect to one of its consequences p is this: make the (kernel of the) result consistent with p; and make
it as extensive a subscheme of A as you can. The first requirement is that of
success; the second requirement is that of minimal mutilation. Moreover,
note that in a scheme, as opposed to a set, mutilation is all the greater the more
proofs or justifications one loses. One would like to be able to conserve as much
as possible of past computational effort, in the form of previously discovered
proofs that ought still to be available after the contraction has been completed.
G L O B A L C O N D I T I O N S O N C O N T R AC T I O N
considerable weight. Cutting all the way back to logic would clearly be a breach
of this injunction in general. We need a further postulate on contraction: one
that will give expression to the requirement of minimal mutilation.
It has been suggested that the following captures this requirement (where
once again we speak of sets rather than schemes):
recovery: If p is implied by A, then whatever is implied by A
is implied by A p, p.
The idea can be attributed to Grdenfors []:
it is reasonable to require that we get all of the beliefs in [our theory] back again after
first contracting and then expanding with respect to the same belief. Formally, this idea is
captured by the axiom [of recovery]. (pp. 9394)
Note that our statement of Recovery has it that contraction with respect to any
consequence p of the set A should admit recovery of all consequences upon
restoration of p to A p. It would not be strong enough, to capture the intended
spirit of minimal mutilation, to require only that contraction with respect to any
belief in A should admit recovery of all consequences of A upon restoration of
that member to the contracted set. For A in general is not logically closed, and
indeed is only finitely developed.
Not much interest attaches, therefore, to the claim of Alchourrn and Makinson [], p. , that recovery holds when applied to irredundant sets of propositions; for their form of recovery is stated with the antecedent If p is already
in A . . . . (For Alchourrn and Makinson, an irredundant set of propositions
is one none of whose members follows from the rest. This is clearly equivalent
to the definition given above.) They are talking about recovery for contraction
with respect to members; we are talking about recovery for contraction with
respect to consequences. In the case of theories, there is no difference between
these two ways of talking, for all consequences are members; but in the case
of belief sets not closed under logical consequence, there is a great difference,
for then there will be consequences that are not members. In the latter case
the really interesting form of recovery is recovery for contraction with respect
W H AT D O W E R E Q U I R E O F A C O N T R AC T I O N ?
93
to consequences, that is, when the proscribed belief newly restored followed
from A, without necessarily being itself a member of A.
Recovery is a seductive idea: it is the supposition that we shall be able give up
so little upon giving up p that, were we to be given just p back again, we would be
able to recover everything (by way of consequence) that we had secured before.
But its attraction is only apparent. Recovery is false; and it fails to give expression
to the requirement of minimal mutilation. Tennant [b] presents a detailed
argument for this conclusion.
Meanwhile, we may draw a methodological conclusion: we stand in need of
a suitable explication of the notion of minimal mutilation, within an account
of contraction that pays proper attention to the structures of reasons within
rational agents belief schemes. Such an explication will be given in Chapter .
The explication is very natural, and is immune to the kinds of counterexamples
to which Recovery succumbs.
94
G L O B A L C O N D I T I O N S O N C O N T R AC T I O N
CHAPTER 4
A Formal Theory of
Contraction
4.1 Introduction
INTRODUCTION
95
The logical aspect of the problem of theory contraction is this: how does one
give up a belief p present in, or entailed by, a system T of beliefs, in such a way
as to ensure that what is left will not give one p back again? The methodological
aspect of the problem is this: how does one give up p with minimal mutilation
of T, so as to obtain a contracted result (call it T p) that is undecided as
between p and its negation p? The computational aspect of the problem is
this: how does one give up p in a methodical, rule-governed way that could be
spelled out as an algorithm? And how complex will such algorithms be?
If we can solve the problem of contraction in general, by dealing with all these
aspects of it, then, to repeat, we shall also obtain a solution to the problem of
belief revision: how does one adopt a belief q that logically conflicts with ones
system T of beliefs, while yet minimally mutilating T? The answer, as already
indicated above, is: first, contract T with respect to its consequence q; then
add q to the result.
There should be a rigorous account of this operation of contraction. Epistemologists and philosophers of science talk of minimally mutilating changes in
the web of belief , but give no constructive account of how to effect them. Hence
we face the challenge of giving a philosophically sound, mathematically precise
account of contraction, programmable on a computer. Automation of the task
of belief-scheme contraction would bring widespread applications.
The account of contraction (hence also of revision) will be normative, not
descriptive: it will reveal what ought to be done by an ideally rational agent
confronted with the need to change its mind. One of the central questions to
be addressed is this:
I understand that giving up p will involve not just giving up p by itself, but will also involve
giving up enough in the way of other beliefs to ensure that what remains is insufficient
to imply p; but does giving up p also involve giving up any belief in the system that wholly
depends on p?
Our answer to this question is affirmative; and this is one of the main features
that sets our account apart from certain others.
The aim in this chapter is to set out an abstract and completely general
framework for the discussion of belief schemes and their contractions; to explicate what exactly is meant by minimal mutilation; to characterize the contraction problem precisely; and to show that this problem, in its simplest form, is
NP-complete.
We treat here of finite belief schemes. Our interest is in how a finite belief
scheme can be contracted with respect to any belief in ithow, that is to say,
one can give the belief up while at the same time inflicting minimal mutilation
96
A F O R M A L T H E O RY O F C O N T R AC T I O N
And we are interested in determining its computational complexity. This chapters main theoretical aim is to show that this decision problem is NP-complete
if one does not worry about the requirement of minimal mutilation. With
that requirement in place, however, the decision problem is prima facie more
difficult. Details will emerge below.
We need to provide formal definitions both for the formulation of the problem and for rigorous proof of the claim that the simplest form of the problem is
NP-complete. It is well known that there are thousands of NP-complete decision
problems. But that fact does not entitle one to dismiss the main result of this
chapter as just another NP-completeness theorem. That would be to underestimate the importance of this particular decision problem for the logic of theory
dynamics and for epistemology at large. This anticipatory defence depends, of
course, on the formal modeling being acceptable as a good first approximation
of the epistemologically important phenomena. But arguing in detail that it
should be so accepted lies beyond the scope of this chapter. (Chapter will
set out the proof of the metatheorem, due to Harvey Friedman, that ensures
that one incurs no loss of theoretical generality by restricting oneself to finite
belief schemes.)
Despite this chapters focus on proving the main theoretical result about NPcompleteness, both the ideas behind our modeling, and the aforementioned
result, need to be easily accessible to the reader who is not a specialist in the
area of belief revision. For the sake of clarity, therefore, we provide illustrations
INTRODUCTION
97
of defined notions along the way. We also make some comments, which are kept
to a minimum, about the motivation behind our modeling of belief schemes
and behind the formal definitions that give it expression. These comments
are expository only, intended to help the reader who is unfamiliar with the
general area.
The reader also needs to know what this chapter is not. It does not specify
any algorithms. (That is done in Chapter .) Nor does it offer any implementations of algorithms. (That is done in Chapter .) One needs to know that
the computational problem is NP-complete in order to be able to assess any
proposed algorithms for solving it. The actual algorithm that is proposed in due
course is a development (in successively less greedy versions) of what was called
the staining algorithm in Tennant [a]. Perhaps now, since our coloring
convention involves whitening the nodes representing the erstwhile beliefs that
are being given up, this algorithm should be re-named the bleaching algorithm.
This chapter is not concerned, either, to provide detailed critiques of, or comparisons with, rival accounts of belief revision and/or contraction. In Chapter
we discuss AGM-theory, JTMS-theory, and Bayesian network theory, so that the
reader can appreciate how the present approach differs from those.
We define all the notions needed for a full understanding of the structure of
dependency networks, that is, finite fragments of theories. These finite fragments
are all that human beings could ever have developed; all that one could in principle be held responsible to take into account when performing a contraction;
and all that any algorithmic contraction method could be applied to. Identifying
successful and minimally mutilating contractions of dependency networks is
non-trivial, especially when non-well-founded justificatory structures are to be
taken into account. One has to draw on significant formal conceptual resources
in order both to explicate the contraction problem and to determine its inherent
computational complexity. The intention here is to introduce those formal conceptual resources by way of sufficient examples and explanatory motivations.
A F O R M A L T H E O RY O F C O N T R AC T I O N
The result holds even when one considers only sentences in conjunctive normal form whose conjuncts are disjunctions of at most three literals. The latter
problem is often referred to as -SAT. SAT is reducible to -SAT in polynomial
time.
There are two kinds of resource that are important for any given class of computational problems. First, there is the amount of memory space that is needed,
at a minimum, by the most efficient algorithm, in the course of executing it in
a computation. Secondly, there is the amount of time that is likewise needed.
These are measured by functions that take as their inputs the lengths of the
problems in question (according to some natural measure).
What does it mean when one says that a computational problem class
(such as deciding whether a given deductive problem admits of proof) is
NP-complete? It means that it can be solved in non-deterministic polynomial
time (that is, it is in NP); and that any other problem class that can be so
solved can be transformed into the one at hand in (deterministic) polynomial
time (that is, it is NP-hard). This in turn raises definitional demands regarding
the notions non-deterministic polynomial time and (deterministic) polynomial
time.
We take the latter notion first. A problem class can be solved in (deterministic)
polynomial time just in case there is a polynomial function f (_) such that any
problem of length n in the given class can be solved in f (n) units of time. These
are also known conventionally as the tractable problems.
Now for the former notion. A problem class can be solved in nondeterministic polynomial time just in case there is an algorithm for solving it,
and a polynomial function f (_), such that for any input problem of length n,
the search tree generated by the branching (choice) points of the algorithm
applied to the problem contains a solution at the end of a branch of length
less than f (n). An intuitive way of thinking of this is that if, in the execution of
the algorithm (which requires us at branching points to make single choices
from a range of possible alternatives) we were so lucky as always to choose
correctlyand thereby solve the problem as quickly as possiblethen indeed
we would do so in no more than f (n) units of time.
That we should be so lucky! . . . P is the class of problems (more exactly:
problem classes) that can be computed in polynomial time. NP is the class of
problem classes that can be computed in non-deterministic polynomial time. It
is an open problem whether P=NP. (This is one of the seven Clay Millennium
problems. The first person to solve it will be eligible for a prize of one million
dollars.) The orthodox conjecture is that P =NP; for no-one has ever provided,
C O M P L E X I T Y O F VA R I O U S D E C I S I O N P R O B L E M S
99
for any one of the hundreds of problem classes now known to be NP-complete,
an algorithm that runs in polynomial time.2
By way of comparison, for the expert:
1. The decision problem for classical propositional logic C is co-NP-complete.
2. Changing classical to intuitionistic makes matters prima facie worse: the decision problem for intuitionistic propositional logic I is PSPACE-complete.3 The
same is true of Johanssons minimal propositional logic4 M.
3. The present authors systems of relevant propositional logic called CR and IR
are of the same complexity as their respective parent systems C and I.5
4. By contrast, the propositional relevance logic R of Anderson and Belnap6 is
undecidable.7
5. Robert Meyers well-known decidable fragment LR8 of R, obtained by dropping
the distributivity axiom, has an awesomely complex decision problem: at best
ESPACE-hard, at worst space-hard in a function that is primitive recursive in the
generalized Ackerman exponential.9
6. For most change operators on propositional knowledge bases T, the question whether T p logically implies q is co-NPNP -complete.10 So, under the
[generally accepted] assumption that the polynomial hierarchy does not collapse,[fn] this means that counterfactual inference is [computationally] harder
than classical inference.11
7. For propositional logics that formalize non-monotonic reasoningReiters
default logic,12 McDermotts and Doyles non-monotonic logic,13 Moores
autoepistemic logic,14 and Mareks and Truszczynskis non-monotonic
N 15 both the problem of testing for the existence of a fixed point (in
the semantics) for a set of premises and the problem of deciding whether
For a lucid introduction to the theory of complexity see Garey and Johnson [].
See Statman []. The proof given there can be adapted to deal also with minimal propositional logic M.
Johansson [].
See Tennant []. Note that Core Logic is now the authors preferred name for IR.
See Anderson and Belnap [].
See Urquhart [].
See Thistlewaite et al. [].
See Urquhart [].
The class NPNP is also referred to as P by theorists of the so-called polynomial hierarchy.
2
For a good explanation of the terminology, see Garey and Johnson [] at pp. .
See Eiter and Gottlob [].
See Reiter [].
See McDermott and Doyle [].
See Moore [].
See Marek and Truszczynski [].
100
A F O R M A L T H E O RY O F C O N T R AC T I O N
So we see that even at the modest level of propositional logic, it is rare to find
a computational logic problem that is as tame as NP-completeas was claimed
in Section ..
R E L AT I V E E N T R E N C H M E N T O F B E L I E F S
101
reject. Evidence must then be assessed, with a view to rejecting the least firmly supported
of the conflicting beliefs.
Since Quine and Ullian countenance varying degrees of support, it is easy for
them to define the relative entrenchment relation in terms of degree comparison. Our account, however, makes the simplifying assumption that belief nodes
that are in (black) have degree of support , and belief nodes that are out
(white) have degree of support . So comparison of degrees of support will not,
for us, yield a relation of relative entrenchment. Nevertheless, such a relation
can still supervene on the other considerations mentioned in the text, and be
useful in making the best choices (for whitening) during a contraction process.
Relative entrenchment is the result of many influences of a mainly pragmatic
nature. The belief a might contribute more than the belief b to the unification
or simplicity of the agents belief scheme. Or the two beliefs might differ in their
degrees of testimonial dependence on the word of others, or the agents estimate
of the reliability of her testimonial sources. There is no implication here that
relative entrenchment could be modeled by suggesting that the agent assigns to
them different degrees of subjective probability or credence. In our modeling
of the agents belief scheme, we can have a relation of relative entrenchment
alongside a discrete on-off (or black-white) assignment to nodes (i.e. beliefs).
Our modeling employs no continuum of values between on and off , as would
be required by a system of probability or credence assignments, which typically
employ the real interval [,] as the space of possible values. For us, there are
only the values and , as it were.
Our modeling also represents the agent as committed to the continuing
reliability of such justificatory steps as she has acknowledged thus far as binding.
The steps always remain the same; and we consider only changes of belief status
on the part of the premise and conclusion nodes of those steps. This means that
any attribution of an inductive source for a step might have to be taken cum
grano salis. For it is a well-known phenomenon with inductive reasoning that
a formerly reliable step a, b, therefore c (whose premises a and b are believed)
can itself be infirmed upon discovery of counterevidence to its conclusion c.
The agent can continue to believe both a and b, but cease to accept the step
from a and b to c. New evidence e, say, could be discovered, and new allegiance
be given to the step a, b, e, therefore not-c. This phenomenon is known as the
non-monotonicity of inductive inference.
In what follows, we do not allow for non-monotonic inferences. We are concerned only with a first and most modest kind of model of belief schemes and
changes to them: those in which the steps remain constant, and only the status
102
A F O R M A L T H E O RY O F C O N T R AC T I O N
Generalizing,
1. If one takes the set X of premises to justify c, and one is giving up c, then one
will have to give up at least one member of X.
2. If in such a situation one prefers some member of X to some other, then (in
the absence of compelling logical reasons to the contrary) one will give up the
latter rather than give up the former.
3. If in such a situation Z W X then one would prefer giving up all of Z to
giving up all of W.
Compare Quine and Ullian [], at p. : Insofar as we are rational, we will drop a belief
when we have tried in vain to find evidence for it. So, if our surrenderings during a contraction
process have deprived a belief of all its sources of support, it too should be a casualty of the process
of contraction.
SOME INTUITIONS
103
4. If in such a situation one believes the proposition b only because one believes
all the propositions in the set X, and one is giving up some members of X, then
one will give up b as well.
The reader will no doubt forgive the author for not having offered this pithy,
perspicuous formulation right away.
of well-foundedness
In Chapters and we shall allay, in the context of deductive theories (or, more
generally: axiomatizable theories in a first-order language), the misgiving that
a computational approach to contraction might suffer from some restriction in
104
A F O R M A L T H E O RY O F C O N T R AC T I O N
its range of application. It does not. We would do well to bear in mind, however, that our treatment of theory contraction will not be limited to deductive
theories, in which justificatory steps from premises to conclusions have to be
logically valid. Rather, we shall take such steps of the form p , . . . , pn ; so, q
to represent justificatory transitions for the (presumably, rational) agent whose
beliefs are in question. In our representation of such an agents belief scheme,
we do not have to be specific about the deductive or inductive status of such
steps. They are simply steps that work as far as the agent is concerned, and that
she will not give up once she has discovered them.
Whatever the nature of justificatory stepsbe they deductive or inductive
not all systems of belief developed by those steps have a well-founded structure
of justifications for their theorems. Let us turn our thoughts for a moment to
theories (or systems of belief) that are not well-founded in their justificatory
structure. According to non-foundationalist, coherentist epistemologists, the
justificatory structures of some systems of belief could involve lines of mutual
logical interdependency, or loops. We shall assume, however, that they could
not involve non-looping, infinitely descending chains of justification among
beliefs. The acquisition or entertaining of any belief takes more than some
fixed amount of time. Hence no finite agent could have progressively acquired,
hence occurrently entertained, each member of such an infinite chain, during
a finite lifetime. The same would hold even of a communal system of beliefs
accumulated by finitely many agents with finite lifetimes. Despite our thus
ruling out of consideration the infinite non-looping case, the non-well-founded
(because looping) case has not in general been excluded. Thus our notion of a
dependency network must be adequate to both the well-founded and the nonwell-founded case.
105
working with a certain set of justificatory steps as given. One works only at the
macro-level.
The macro-level is a level of abstraction one up from the assignment of logical
forms to sentences and the negotiation of deductive passages among them.
In the interregna between contractions and revisionsperiods of inquiry that
might also include expansions of the belief schemethe agent will be discovering (and recording) new justificatory steps by working at the micro-level. But
as soon as one is confronted with a need to contract the scheme, one can ascend
to the higher, or macro-, level and regard the sentences in ones belief scheme as
just so many internally featureless nodes within a directed graph that contains
all the justificatory steps discovered thus far. Cf. Doyle [], at p. , when
treating his truth maintenance system:
dependency-directed backtrackingonly operates on the reasons for beliefs, not on
the form or content of the beliefs.
In our terminology, de Kleers problem solver works at the micro-level, and his
assumption-based TMS works at the macro-level.
In this way also the method of contraction is absolutely indifferent to what
underlying logic might be employed at the micro-level. It is invariant across
choice of logic. This higher level of abstraction affords the prospect of graphtheoretic classifications of belief schemes according to their behavior under the
contraction operation. Intelligent systems could then recognize a new contraction problem as graph-theoretically isomorphic to an earlier one that might
have involved completely different sentences, with completely different internal
logical structures, and concerning a completely different subject matter. The
gain from such abstraction is a wider range of applicability, without an eye to
distractingbecause irrelevantinternal details.
Our dependency networks are made up of steps connecting nodes via inference strokes.
CONSISTENCY CONSTRAINT.
dependency network.
106
A F O R M A L T H E O RY O F C O N T R AC T I O N
The consistency constraint does not, however, prevent the agent from being
apprised of steps of the form p , . . . , pn |. Such steps can play an important
role, which will be explained in due course. They cannot, however, appear in the
agents actual belief scheme (i.e. in the black part of it), since that scheme is supposed to capture the agents beliefs. And, since the agent is assumed to be a logical paragon, she will not believe absurdity (). It will always be white for her.
As already observed, a given step (even one with conclusion ) finds its way
into an agents repertoire of known steps through some process of discovery or
recognition on the agents part, a process that may well turn on details of the
internal logical structure of the sentences that label the nodes in question. But
the step itself, once adopted, represents only the fact (at a grosser level) that,
as far as this agent is concerned, one may justify belief in the conclusion node
of a step (when that conclusion is not ) by appeal to belief in all its premise
nodes. That is the ground zero, or lowest substratum of relevant facts, that will
be available to guide the agent (or guide us, on the agents behalf) through any
subsequent process of contraction (or revision).
For these reasons, we adopt here the following methodological maxim:
MACRO-STRUCTURAL CONSTRAINT.
107
belief schemes
In our formal modeling, we are representing any rational agents finitary belief
scheme as consisting of finitely many sentences (or nodes), arranged in finitely
many steps. The collection of such steps is a dependency network. (Note that
we have yet to state a precise formal definition of the notion of dependency
network. Everything we have said thus far about dependency networks has been
designed to convey the intended notion, so that the formal definition, when it
is given, will make good sense.) Every node figuring as a belief of the agent
must stand as the conclusion of some stepeven if this is a step of the form x|x,
indicating that the belief x is, as far as the agent is concerned, directly warranted.
Beliefs that are only indirectly warranted within the agents belief scheme will
stand as conclusions of steps whose premises are in the scheme in question.
4.7.1 Nodes
Nodesfeatureless and unstructuredare to represent particular beliefs within
the belief scheme of a rational agent. They represent, as it were, specific propositions that the agent might or might not believe. Nodes are the Urelemente
of which we treat. Everything else that we shall be talking about will be built,
Bourbaki-style, out of nodes by forming hereditarily finite sets. In this project
of theoretical modeling, one must be careful not to take certain convenient
artifacts of the model as representing would-be, but absent, features of the
situation being modeled. We shall have occasion to remind the reader of the
need for caution on this score when we see how we treat initial beliefs in our
formal modeling.
108
A F O R M A L T H E O RY O F C O N T R AC T I O N
U is called the premise set of the step (U, x); x is called the conclusion of the step
(U, x).
Observation 1 For a given node x there are only two possible forms of step with
x as conclusion:
y , . . . , yn |x where each yi is distinct from x ( i n); and x|x.
We shall call these transitional and initial steps, respectively.
Every step is either initial or transitional. No step is both initial and transitional.
The reader is advised not to confuse our use of | here with other widespread
uses of that symbol. The latter include the use on which it is read as such that
within a set-abstract like {x | F(x)}; as well as its use in logic programming,
where it means given, as in procedural rules of the form
goal | subgoal , . . . , subgoaln .
Our use of | might look like a single turnstile shorn of its horizontal stroke:
to be sure, premises occur to its left, and a conclusion occurs to its right. But
what is designated overall is a justificatory step (for the agent). No claim of
logical deducibility is made or implied by the use of |. For, first, x , . . . , xn | y
is a complex noun phrase, meaning the step whose premises are the nodes
x , . . . , xn and whose conclusion is the node y. By contrast, , . . . , n
is a statement, to the effect that there is a proof of the sentential conclusion
from the sentential premises , . . . , n . Secondly, the nodes are featureless, and
devoid of logical structure. That is not to say, however, that an agent might not
know of a certain step because of an actual deductive proof in her possession,
leading from the sentential premises corresponding to the premise-nodes of the
step in question to the sentential conclusion corresponding to its conclusion
node. It is just to say that when such a step is represented as a transition, not
among sentences, but among featureless nodes, such deductive provenance can
be, and is, suppressed. This is in the interests of revealing only such justificatory
macro-structure within a belief scheme as is relevant for the purposes of contraction. This structure resides at a grosser and more abstract level than that of
the detailed logical forms of sentences themselves.
Definition 2 Let T be a set of steps, called T-entries. For any step (U, x) in T, U
is said to be x-generating or (when x = ) x-justifying (in T).
N E W N OT I O N S F O R M O D E L I N G O F B E L I E F S C H E M E S
109
Put another way: steps in a rational agents repertoire require all their premises.
If an agent comes to realize, of a given step in her repertoire, that she can reach
its conclusion by using some but not all of the premises of that step, then she
will discard the weaker step in favor of the stronger step that uses the weaker
(reduced) set of premises.
From now on we assume that all (our formal modelings of) a rational agents
belief scheme, and repertoire of known steps, satisfy the perfection constraint.
A step whose conclusion is distinct from each of its premises is called transitional. In our modeling, a transitional step is understood as being something
of which the agent is aware, or which the agent has acknowledged, as rationally
compelling for her. Hence it is a normative requirement that the agents system
of beliefs should be closed with respect to all such steps. One is, in effect, saying
to the agent Since you yourself acknowledge the step from x , . . . , xn to y as
110
A F O R M A L T H E O RY O F C O N T R AC T I O N
compelling, and since you believe each of x , . . . , xn , you should also believe y.
Note that such obligations are agent-relative, since the nodes are featureless.
We are not adverting to the logical structures of the sentences expressing the
beliefs x , . . . , xn and y, and insisting that the agent should recognize the step
as logically valid. Rather, it is taken as a given that the agent herself, for whatever reason, has already acknowledged the step in question as compelling. Her
obligation is now only that of taking the step to y should she ever come to believe
all of x , . . . , xn .
Although nodes are basic, in that steps are built up out of them, we shall find
that steps themselves are the most convenient data type with which to work.
Note that an agent can know of a step without being committed to any of its
nodes, as beliefs. But if an agent knows of a step all of whose premise nodes she
believes, then we take her to be committed to believing the conclusion of that
step as well.
Given the perfection constraint, any step with more than one premise
will be transitional. For steps with exactly one premise, there are two possibilities: those where the conclusion is distinct from the premise and hence are
transitional; and those where the conclusion is identical to the premise (and
hence are non-transitional).
In the latter case, we acquire a convenient way of representing an initial belief:
one that is believed outright, in the sense that (as far as the agent is concerned)
it requires no further support from any other beliefs that the agent happens to
hold. The operational meaning of an initial step, i.e. one of the form p|p, is to be
that p is believed outright, without justificatory appeal to any other beliefs. The
diagrammatic representation of such an initial step is as follows (as explained
in Chapter ):
N E W N OT I O N S F O R M O D E L I N G O F B E L I E F S C H E M E S
111
112
A F O R M A L T H E O RY O F C O N T R AC T I O N
Recall that steps fall into two disjoint kinds. There are the initial steps, whose
reflexive-seeming Bourbakization p|p we have explained above; and there are
the transitional steps, that take one from certain premise nodes to another,
quite distinct, conclusion node. We repeat, however, that it is compatible with
ps being an initial belief that p stand also as the conclusion of some other,
transitional, step, all of whose premises the agent believes. Such would be the
case, for example, if p were a simple observational belief that the agent had
acquired in the usual way, but for which the agent also had theoretical support
by way of, say, prediction within a higher-level theory on the basis of yet other
observational beliefs.
Any rational agent who has a system of beliefs might treat some of these only
as initial, that is to say, might believe them outright, without offering yet other
beliefs in support of them. In all other cases, a belief will have support within
the agents system. That is, the belief in question will stand as the conclusion of
at least one transitional step that involves at least one premise, and all of whose
premises are in turn believed. Naturally, a belief can be over-determined, by
enjoying support from many distinct (though not necessarily disjoint) premise
sets, each of which is sufficient (from the agents point of view) to justify the
belief in question. (This is why we had to be so careful in formulating the
generalized intuition (*) in Section ..) This is what makes the contraction
problem, in general, so interesting. For, in giving up such a belief p, the agent
will have to give up at least one premise within the premise set of each step that
has p as its conclusion.
We argued in Section . that we can omit from a rational agents belief
network any nodes labeled by sentences that she knows to be logically true. We
do, however, continue to cater for the possibility that a logical falsehood might
pop up within a belief scheme. It may take time for the agent to discover that
absurdity lurks in her belief schemetwo contrary beliefs p and qbut as soon
as she is aware of that fact she must contract with respect to the conclusion
of the newly recruited step
p
whose presence, now, in her repertoire of known steps is precisely what her
awareness of that fact consists in. (Note that we do not say presence, now, in
her belief scheme, since it is required of any step within a belief scheme that its
N E W N OT I O N S F O R M O D E L I N G O F B E L I E F S C H E M E S
113
premises and conclusion be believed.) Note that in many cases q will be of the
form p, the explicit contradictory of p. But we are allowing here also for pairs
of contraries such as It is nutritious and It is poisonous. 21
Our modest consistency requirement on a rational agent is not that her
beliefs should always be consistent tout court, but only that they should be
consistent according to the steps of which she is apprised. 22 As soon as she
becomes apprised of a new step whose conclusion is , the use of which renders
her current beliefs jointly inconsistent, she must undertake the necessary process of contraction (with respect to ) in order to get rid of the inconsistency.
The rationality maxim in question is therefore Always be prepared to contract
with respect to !, rather than the unrealistically demanding Do not ever form
beliefs from which can be derived!unless, of course, she already knows the
steps by means of which can be derived. The way this would be represented
in our diagrams is to insist that the absurdity node and the inference stroke
immediately above it always be white:
p
Here we color the nodes p and q gray, in order to emphasize the obligatory
whiteness of what is below them.
The operational meaning of a transitional step, i.e. one of the form
y , . . . , yn |x, is that for the agent, the premises y , . . . , yn would justify the
conclusion x. If moreover she believes each of the premises y , . . . , yn , then she
does believe the conclusion x, since she is a logical paragon. The presence of
this step within the agents belief scheme (let us call that scheme B) represents
the agent both as believing its premises y , . . . , yn and hence also as believing its
conclusion x. Knowledge of the step is one thing; believing all its premises, and
hence believing its conclusion, is quite another. In Chapter we used black to
indicate belief:
See Tennant [], where a case is made for logical contrariety being more fundamental than
the notion of a propositions contradictory.
Even this modest a requirement puts one in disagreement with Foley [].
114
A F O R M A L T H E O RY O F C O N T R AC T I O N
...
y1
yn
Once a transitional step is known to the agent, it remains a permanent conduit, and represents a commitment on her part. So, whatever the patterns of
colorationhere is the same step but with a different pattern of coloration:
...
y1
yn
this little bit of the jungle gym remains as a step known to the agent.
This too is a rather modest requirement of rationality. If the agent were ever
to give up the belief x, then she would have to give up also at least one of
the beliefs y , . . . , yn . (The last diagram represents her as having given up at
least yn .) Moreover, if thereafter the agent were ever willing once again to assert
all the premises y , . . . , yn of the step, then she would be willing also to assert its
conclusion x. Such relative justification is internal to the agent, and we do not
inquire after its legitimacy. We can prescind from all questions about the nature
of the justification, about the status of such intermediate inferences as she may
have employed, and so on. The agent sets her own standards of justification; but
then we take her as beholden to them.
115
Every step in C \ B would, on our coloring convention, have its inference stroke
white, even if its conclusion is not . And any step whose inference stroke is
white must have that inference stroke positioned in the fringe C \ B. The conclusion of such a step, howeverprovided it is not could still be black, by virtue
of support received from elsewhere, within B. And every black conclusion node
lies within B. For example:
116
A F O R M A L T H E O RY O F C O N T R AC T I O N
The picture, then, is that the transitional part of the agents belief network B
is (most likely, properly) contained within the set C of transitional steps known
to her. In the case of proper containment, C \ B is non-empty. So there is a step
y , . . . , yn |x in C that is not in B. This means that the agent does not believe all
of the premises y , . . . , yn (so at least one of them would be white, along with
the steps inference stroke); otherwise, the step would be in B.
Moreover, C can contain steps with conclusion . Since B is consistent modulo C, this means that given any such step, not all of its premises can be in B.
Consider such a step p, q|. If its premise p is in B, then its other premise q
must lie outside B:
..
.
B
C
N E W N OT I O N S F O R M O D E L I N G O F B E L I E F S C H E M E S
117
A F O R M A L T H E O RY O F C O N T R AC T I O N
119
3. {({a1 }, a1 ),, ({an }, an )}, the dependency network consisting of the selfjustifying statements a1 , . . . , an . We can also write this as
| a1 , . . . , | an .
4. {({a}, a), ({b}, b), ({a, b}, c)}, the dependency network consisting of the two selfjustifying statements a, b and their immediate consequence c. We can also write
this as
| a,
| b,
a, b | c.
5. {({a}, b), ({b}, a)}, the dependency network consisting of the two mutually justifying statements a, b (making up a shortest possible loop). We can also write
this as
a | b,
b | a.
6. {({a1 }, a2 ), ({a2 }, a3 ),, ({an }, a1 )} (n 2), the dependency network consisting of the statements a1 , . . . , an arranged in a longer justificatory loop. We can
also write this as
a1 |a2 ,
a2 |a3 , . . . ,
an |a1 .
7. {({a}, a), ({c}, b), ({a, b}, c)}, the dependency network consisting of the selfjustifying statement a, the statement b depending on c, and the statement c
depending on a and b. We can also write this as
| a,
c|b,
a, b|c.
8. {({a}, a), ({b}, b), ({a}, c), ({b}, c)}, the dependency network consisting of the selfjustifiying statements a and b, and the statement c, which depends on a and also
depends on b. That is, each of a and b by itself is sufficient justification for c. We
can also write this as
| a,
| b,
a | c,
b | c.
A F O R M A L T H E O RY O F C O N T R AC T I O N
or wholly black:
or
c
121
122
A F O R M A L T H E O RY O F C O N T R AC T I O N
N E W N OT I O N S F O R M O D E L I N G O F B E L I E F S C H E M E S
123
4.8 Closure
In the course of the agents investigations, she can come to accept or acknowledge many transitional steps of the form
y , . . . , yn |x
124
A F O R M A L T H E O RY O F C O N T R AC T I O N
while yet failing to believe all of y , . . . , yn . The belief x might fail to enjoy any
support, within the agents belief scheme, from sufficient sets of beliefs (from the
agents point of view). In such situations the known steps with x as conclusion
are not useless. For, if ever the agent were to come to believe all the premises
of any such step, its conclusion x would thereby acquire rational support (from
her point of view). This could happen, for example, if the agent at first believed
all of y , . . . , yn but did not believe yn ; and subsequently came to believe yn .
At that point, since the step
y , . . . , yn |x
is known to the agent, she would be rationally obliged to adopt the belief x. That
is, she would have to close her beliefs under the step in question.
Clearly, such obligatory operations are iterable, with respect to all known
steps all of whose premises have come to be believed. But the iterations will
only ever be finite in number. This is because only finitely many steps are known
to the agent, and each step involves only finitely many premises. The need for
these iterable operations arises only when the agent adopts one or more new
beliefs.
Now it is important not to fall prey to the confused thought that the agent
may know of schematic rules that correspond to an infinity of steps. We reiterate
the point that our nodes are featureless, and their labels are not to be thought
of as endowed with any essentially logical structure. Each node represents a
particular belief, as given by a particular proposition, expressible by a particular
sentence. No agent will ever have justified to herself more than finitely many
such beliefs. Nor will she ever have performed, in her thinking, more than
finitely many steps. Indeed, at no time in the future of humankind will the
communal belief scheme, and its repertoire of known steps, even assuming
perfect memory, ever contain more than finitely many beliefs, or involve more
than finitely many steps. We call this the finitary predicament. Fortunately, it is
what makes a computational account of contraction and revision possible! To
put the matter succinctly: for the purposes of belief representation, schemata are
irrelevant; only their instances count. And at any point in time, there will have
been only finitely many instantiations.
We shall call the outcome of the iterable operations described above the
universal closure of the agents belief scheme, within the set of steps known to
the agent, and with respect to the agents set of newly adopted beliefs. We are
assuming that in this context there is no possibility of confusing this kind of
universal closure with the logicians syntactic notion of the universal closure of
CLOSURE
125
a formula with free variables, which closure is obtained by prefixing the formula
with universal quantifiers to bind its free variables. Universal Closure is called
propagate inness by JTMS-theorists; see Forbus and de Kleer [].
The situation can be schematized as follows. Let C be the finite set of all
transitional steps known to the agent. Every transitional step in the agents
current belief scheme B (which itself is finite) will be in C; but in general there
could be (transitional) steps in C that are not in B. Moreover, B can contain
initial steps, none of which is in C. Note that in the belief scheme B, every node
of every step is believed, whether it be a premise or a conclusion of the step in
question.
By way of illustration, take the current belief scheme B to consist of the steps
a|a (initial)
e|e (initial)
e|h (transitional)
and let the set C of transitional steps known to the agent be
e|h
a, b|d
c, d|f
d, e|g.
A F O R M A L T H E O RY O F C O N T R AC T I O N
Let P now be a finite set of nodes not in B. The nodes in P are to be understood as newly adopted initial beliefs. If, upon adopting all the new beliefs
in P, the agent finds that she now believes all the premises of some step in
C \ B, then she is obliged to close her beliefs under that step, thereby possibly
acquiring a further new belief (namely, the conclusion of the step in question).
The eventual outcome of iterating the operation will be the universal closure
discussed above. In full detail, it is the universal closure of B within C with respect
to Pabbreviated [B; P]C .
Let us illustrate the process of universal closure with the diagram above.
Suppose we take P to be {b}. Then the universal closure [B; P]C is easily seen
to be the black part of the following diagram:
Dependency networks can in general contain subnetworks. The task of contracting a network T with respect to one of its nodes p will be the task of finding
a suitable subnetwork: one that does not contain the node p, but that satisfies
CLOSURE
127
the rational closure requirement (with respect to the transitional steps in T).
A further requirement would be that the contraction is, in some suitable sense
(to be explicated below) maximal. That is to say, the subnetwork R in question
must be the outcome of a minimally mutilating process of contraction of the
original network T. And the subnetwork R must itself be closed with respect to
the transitional steps in T.
Why require such closure? Because the subnetwork R is supposed to represented a new rational reflective equilibrium for the agent. She is still rationally
beholden to close her surviving beliefs under all the steps known to her. Even
though the status of nodes may change, the steps remain, as conduits of commitment that will continue to compel the agent from their premises to their
conclusions. The agent will have achieved reflective equilibrium after carrying
out the contraction process in question only if that equilibrium represents her
as having made all the justificatory transitions that she knows about. The justificatory transitions in question are exactly the transitional steps in T. We must
remember that in giving up a step, the agent is not losing, or eschewing, or
no longer claiming to be aware of, the justificatory transition that it represents;
rather, she is simply reaching a state in which she no longer believes all its
premises. The step itself, however, is still there for her to use; and use it she
ought. Her new system of beliefs must be closed; it must be a T-closed subsystem
of her original system T.
We shall explicate the notion of maximality below. It is important to realize
that in general a network T might contain more than one maximal closed subnetwork not containing p. That is, the operation of contracting T with respect
to p can in general yield more than one result. This will be reflected in the nondeterministic character of any algorithms that one might devise for performing
contractions.
The complaint has been raised, by theorists immersed in the AGM-tradition,
that any procedure for contraction should tell one exactly which contraction to
adopt. But this is not, in general, possible! Contraction is not functional. That is,
contraction is not an operation with uniquely defined values. No matter what
extra information one avails oneself of, in the way of relative entrenchment of
beliefs (say), there is always the logical and mathematical possibility that there
is more than one equally good contraction of a given system of beliefs with respect
to any chosen one of the beliefs therein.24 Our account has the great virtue of
accepting this possibility (which tends to be the rule rather than the exception),
Tennant [a] offers a friendly amendment to AGM-theory, allowing for multiplicity of
contractions and revisions.
128
A F O R M A L T H E O RY O F C O N T R AC T I O N
and turning to the task of how one accounts for all the eligible contractions
in the multiple-valued case. Indeed, the contraction algorithm(s) described in
Chapter systematically generate all the eligible contractions; and in the Prolog
implementation that we favor, this is achieved by back tracking from one good
solution to another, until all the equally good solutions have been enumerated.
Algorithm 4.8.1: U n i v e r s a l C lo s u r e C, B , P
Steps C
Closure B the set of initial steps involving members of P
while some step in Steps but not in Closure has each of its premises standing
as the conclusion of a step in Closure
The algorithm must terminate, since C is finite. We say that it determines the
universal closure (of B within C with respect to P) because of the word each in
the while condition.
Universal closure is exceptionally simple. Indeed, even the prima facie more
difficult problem of deciding whether a given step can be derived using finitely
many given steps is solvable in linear time. In this connection, see Dowling and
Gallier []. The problem is exactly that of querying a Prolog program with
finitely many propositional clauses. A conditional Prolog clause such as a :- b, c.
is precisely the transitional step b, c|a. An unconditional Prolog clause a. is
precisely the initial step a|a. To find whether, say, the step p, q, r|s follows from
finitely many other steps, simply form the Prolog program consisting of the
clauses corresponding to the latter steps, adjoin the Prolog clauses p., q. and r.
to the program, and query the expanded program with s.
CLOSURE
129
Remember that T-closure involves only the application of steps that are never
called into question during the contraction process.
The honesty condition says that one should not be able to close Tp any
further by means of T-steps. Thus a contraction contains all nodes known to be
consequences of any nodes that it contains.
The success condition says that Tp should not contain any step of the form
(U, p). Put another way, p should not be in |T p| = |[Tp]T |.
Definition 12 A contraction R of T with respect to p is minimally mutilating just
in case it satisfies the further condition
3. (MINIMAL MUTILATION) R is inclusion-maximal with regard to (1) and (2).
The minimal mutilation condition ensures that a contraction is as economical as possible, in the sense that one gives up as little as possible in the
way of T-steps when passing from T to Tp. Here, giving up a step means
changing its inference stroke and at least one of its premises from black to white.
The step itself, of course, remains, relegated to C \ B, for possible future use in
expansions, as we have already explained. The minimal mutilation condition
seeks to conserve the outputs of past computational effort in developing T.
As de Kleer [], p. , put it: Thus inferences, once made, need not be
repeated . We are therefore requiring a minimally mutilating contraction to
be a maximally non-(p-generating), T-closed subnetwork of T. So a minimally
mutilating contraction Tp will contain all those T-steps compatible with the
requirement that p not be in |T p |.
This is the intuitively clear, mathematically precise explication of minimal
mutilation that we have been looking for. No more hand-waving will be needed.
And we confidently conjecture that no intuitively compelling counterexamples
to it will be found.
130
A F O R M A L T H E O RY O F C O N T R AC T I O N
131
A F O R M A L T H E O RY O F C O N T R AC T I O N
C O N S I D E R AT I O N S O F C O M P L E X I T Y
133
the network, at least one of its premises. (Naturally, we seek to minimize this
subset in order to maximize the resulting contraction.) The isomorphism with
graphs is easy to see: the (unordered) pair-sets of premises (of the steps) map
to edges of a graph, which are (unordered) pair-sets of vertices. Ones choice of
a set of premise nodes to be removed (at least one from each pair-set) in order
for the common conclusion of all the two-premise steps to have no justification
then corresponds to a choice of a set of vertices in the graph that covers each
edge (by containing at least one vertex from each edge, i.e. from each pair-set of
vertices).
QED
The classification of problem classes in complexity theory requires that one
state whether one is dealing with a decision problem, or with a (functional)
search problem. For given inputs x, and effectively decidable relation L(x, y), the
decision problem takes the form of asking whether yL(x, y), and the outputs
of any algorithm for solving such a problem are accordingly Yes/No; whereas
the corresponding search problem requires as output either a computed value
such that L(x, ) (if such a value exists), or the answer No if no such value
exists. Whereas the decision problems are assigned to the classes P, NP, etc., the
corresponding search problems are assigned to the classes known as FP, FNP,
etc., where the F stands for functional. This terminology is now standard, even
though there is no guarantee of uniqueness on the part of any values such that
L(x, ).
When a decision problem is hard, then the corresponding search problem is
at least as hard. By Theorem , the problem of searching for a contractionand
so, a fortiori, for a minimally mutilating contractionis at least FNP-hard.
Unlike the problem of deciding whether there is a contraction of a certain size
a contraction that is required to satisfy only conditions () and (), so that this
decision problem is in NPthe problem of searching for a minimally mutilating
contractionone satisfying in addition condition ()is not necessarily in
FNP. 27
When minimally mutilating contractions are the objects of interest, we pose
the problem as a search problem rather than as a decision problem. This is
because minimally mutilating contractions are inclusion-maximal; so it is trivial
that a Yes [No] answer to the question Is there a (simple) contraction of size at
least such-and-such? implies the same Yes [No] answer to the question Is there
a minimally mutilating contraction of size at least such-and-such?.
134
A F O R M A L T H E O RY O F C O N T R AC T I O N
C O N S I D E R AT I O N S O F C O M P L E X I T Y
135
calls to an FNP-oracle. What we can actually show (thereby securing membership in the subclass FPFNP[O(n)] ) is that the problem can be solved by making
no more than a linear number of calls to an FNP-oracle. The FNP-oracle will be
called upon here to solve instances of the simple decision problem above.
Here is the search strategy.
First, ask the FNP-oracle to look for R with inputs T, p and k = |T| .
If some R is found satisfying condition (i), then |R| = |T| and condition (ii) is automatically satisfied, since R itself is as large a solution as can be.
If with inputs T, p and k = |T| no R is found satisfying condition (i),
then ask the FNP-oracle to look for R with inputs T, p and k = |T| .
If some R is found satisfying (i), then once again condition (ii) is automatically satisfied. This is because failure to satisfy condition (ii) means that there
is some U such that U is a T-closed subnetwork of T not containing p, of size
at least |T| . But the existence of such U would have registered earlier as
success in the search for R satisfying condition (i) with k = |T| .
If with inputs T, p and k = |T| no R is found satisfying condition (i),
then ask the FNP-oracle to look for R with inputs T, p and k = |T| .
Clearly, since we know that a satisfactory object of our search does exist, we
shall make at most |T| calls to the FNP-oracle before we find some R such that
conditions (i) and (ii) hold.
QED
As shown by Georg Gottlob (personal communication), Theorem can be
improved.
Let us abbreviate the conjunction of (i) and (ii) as L(T, p, k, R).
Theorem 4. (Gottlob) The problem of searching for a minimally mutilating contraction of a given minimum size is in FPFNP[O(log(n))] , since it can be solved in
polynomial time with at most a logarithmic number of calls to an FNP-oracle.
Proof. The problem of searching for a minimally mutilating contraction is to
be understood as follows: given as inputs a dependency network T, a node p
therein, and an integer k, find some R such that L(T, p, k, R).
The search algorithm first determines, via binary search with O(log(n)) calls
to an NP-oracle, the maximum number m such that
m < |T| R L(T, p, m, R).
(Here it actually suffices to make the intended calls to a Yes/No oracle, i.e. an
oracle for a decision problem.)
136
A F O R M A L T H E O RY O F C O N T R AC T I O N
|T|
, R).
|T|
, R);
|T|
, R),
and so on. Once the maximum such m has been determined, we call finally
on a functional (FNP) oracle to find some R such that
L(T, p, m, R).
Since m has already been determined as a maximum in this regard, any solution
R that the FNP-oracle returns will be such that
(i) R is a T-closed subnetwork of T not containing p, of size exactly m; and
(ii) there is no U such that R U T and U is a T-closed subnetwork of T not
containing p
This procedure uses, altogether, no more than log(|T|) + oracle calls. Thus
the algorithm runs in polynomial time with O(log(n)) oracle calls.
QED
Theorem is an improvement that consists in making at most a logarithmic
number of calls to an oracle, rather than a linear or polynomial number of calls.
As Gottlob has also pointed out, there is a way to emulate a call to an FNP-oracle
by making a polynomial number of calls to a Yes/No NP-oracle. This enables
one to re-cast the proof of Theorem so as to make it a proof of the conclusion
that the search problem for a minimally mutilating contraction is in FPNP . But
this does not appear to admit of further improvement to FPNP[O(log(n))] . The
latter class is widely believed to be properly included in FPFNP[O(log(n))] , which
is widely believed to be properly included in FP , which is equal by definition
to FPFNP , which in turn is equal (by the argument for search-oracle imitation
in polynomial time by a bit-oracle) to FPNP .
We leave as an open problem that of determining whether the minimally
mutilating contraction problem is complete for FPFNP[O(log(n))] i.e. whether
it is FPFNP[O(log(n))] -hard.
C O N S I D E R AT I O N S O F C O M P L E X I T Y
137
One should not be at all surprised or dismayed to find that the problem
of belief contraction in its simplest form is intractable in the sense of being
NP-complete. The interesting challenge was how to conceive of the problem in
a manner sufficiently amenable to complexity analysis in the first place. It is
in this spirit that we have conducted our investigations. The result has been
encouraging. We have discovered that the simplest form of the problem of
theory contraction, conceived in this way, is of the lowest complexity that one
could have expected it to be, given the considerations laid out in Section ..
Moreover, theorists of contraction who object (misguidedly, in our view) to
representing belief schemes in our finitary fashion now have to contend with the
prospect that their own contraction decision problems (in so far as they might
be able to define any, on their approaches) will have much worse lower bounds
of computational complexity. If NP-completeness, or a region quite low down
in the second level of the polynomial hierarchy is bad news for the dependencynetwork theorist, then it will be just as bad, if not worse, for AGM-theorists
if ever they produce any computational implementations of their contraction
functions.
4.11 Conclusion
We have laid the theoretical foundation for a general computational account
of theory contraction, by conceiving of belief schemes in an appropropriately
finitary wayas dependency networks. The treatment marks a retreat from
the infinitistic objects of AGM-theory, and a return to the level of description
favored by earlier TMS-theorizers, in which all that is important (for the purposes of contraction) is the pattern of justificatory connections thus far disclosed among the agents various beliefs, and not the internal logical structure of
the sentences involved, nor the underlying logic that would exploit such structure. By paying too much attention to the latter features, AGM-theory had made
the contraction problem non-computable. We have regained computability by
abstracting at the appropriate level. We can also supply principled reasons why
one can limit oneself to finitary objects without any loss of theoretical generality.
Our formal modeling is arguably more precise and logically cleaner than
earlier TMS-theorizing. It exactly captures the intuitions involved in the various
counterexamples that have been advanced against AGM-theorys postulate of
recovery. It allows for the definition of contraction in such a way that minimal
mutilation is captured by inclusion-maximality. That mathematical definition
138
A F O R M A L T H E O RY O F C O N T R AC T I O N
allows one to characterize the decision problem for contraction of belief networks as NP-complete (if one does not require minimal mutilation) and the
search problem as no worse than FPFNP[O(log(n))] (if one does require minimal
mutilation).
So the requirement of minimal mutilation really elevates the contraction decision problem to a prima facie strictly higher level of computational
complexitythough no further than (quite low down in) the second level of
the polynomial hierarchy. One might take this fact to provide at least a partial
explanation of why the notion of minimal mutilation appears to have resisted
precise characterization for so long. One would also expect rational agents to
have developed heuristics that, while aimed at producing contractions without
checking them for minimal mutilation, nevertheless tend to produce, at a decent
rate, contractions that do happen to be either minimally mutilating, or almost
minimally mutilating.
The following chapter explores some of the heuristics that might be employed
in algorithms designed to guide contraction processes that are reasonably minimally (even if not absolutely minimally) mutilating.
CONCLUSION
139
CHAPTER 5
Specification of a
Contraction Algorithm
140
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
the coin of highest value that does not exceed the outstanding amount. With the
US currency of quarters, dimes, nickels and cents, this algorithm always yields
the best solution. Indeed, it will do so with any currency each of whose coinvalues is at least twice that of the next lower coin-value (if there is one).
The algorithm is called a greedy onenot because it involves money, but
because it seeks to take as little time and to consume as few resources as possible
in the course of producing a solution. At each stage, the algorithm dictates a
local choice, which can be made very efficiently. The sequence of those local
choices produces, happily, a globally optimal solution. Greedy can be good;
indeed, best possible.
Whether a greedy algorithm produces a globally optimal solution depends
however, on the nature of the problem being solved. It is easy to imagine coin
denominations other than those of the US currency, for which the greedy algorithm just described would produce suboptimal solutions. Suppose some weird
cult were to gain control of the nation, and outlaw quarters and nickels, while
introducing also the septickel, a coin worth cents (to remind us, say, of the
seven deadly sins). If one wished to give cents in change, the greedy algorithm
would serve up a dime and then four penniesfive coins in all. It would miss
altogether the optimal solution of two septickels.
The advantage of a greedy algorithm is that it operates fastor at least, as fast
as is permitted by the local considerations that enter into the determination
of each successive choice. In general a greedy algorithm makes its successive
local choices very efficiently, and uniquely, hence offers up an overall solution
quite quickly. The emphasis on unique choice is important: with the changegiving problem, the next coin to be chosen is uniquely determined by the
amount outstanding. There is no branching in the decision tree. The decision tree is really a decision chaina steady march from one choice to the
next. For this reason the algorithm is called deterministic. The big question,
however, is whether the problem being solved is one for which suboptimal
solutions can be tolerated. The suboptimality of the solution has to be arbitraged
against the extra computing costs that would be involved in finding an optimal
solution.
In the field of Artificial Intelligence, the technique of providing an efficient
algorithm that produces good enough solutions to the problem at hand is called
satisficing. The idea was introduced by the Nobel laureate Herbert Simon. The
notion of satisficing has proved useful also in the field of evolutionary biology.
Certain adaptive modifications of body-plan, for example, might be absolutely
G R E E DY A L G O R I T H M S
141
ideal for a given species under certain selective pressures. Yet it may be that
too much genetic change would be called for in order to achieve the best
possible modification; and it may be that intermediate stages through which
the changing body-plan would have to pass in order to achieve the best plan
possible would actually be very suboptimal. (The species would have to descend
into a valley in the landscape of fitness values before being able to scale the
neighboring, much higher peak. One has to think here of the underlying plane
for this metaphor as made up of points corresponding to the phenotypes involving certain body-plans.) So natural selection will satisfice by producing as
much change in that direction as possible, by using up the available genetic
variability for the re-shaping that improves overall fitness. It will drive the
species higher on that mountain in the landscape of fitness values on which
it is currently located. And whether the species reaches the absolute summit of
that mountain depends again on the costs involved in getting there. It may do
very well by getting sufficiently close to that local peak; and too much change
might be called for in order to get it to occupy the peak itself. In the parlance
often employed, it is enough to get within an epsilon of where, ideally, one
would wish to be. For further discussion, see von Schilcher and Tennant [],
pp. .
The way extra computing costs can be involved is that one may wish to look
ahead, further down the decision tree, to consider branching points at which
alternative choices may be available for continued construction of a solution.
One looks ahead in order to tell whether a choice now being contemplated is
one that could have inevitable and bad knock-on effects further down the line.
In general, a non-deterministic algorithm is one that involves exploring all the
possible branches within a decision tree, and evaluating the solutions that would
be obtained upon reaching the end of any one branch. The more one takes into
account in looking ahead, the less greedy ones algorithm becomes. As we saw
in Section .., the source of non-determinism in the contraction problem
is the need to make a choice of a premise node to whiten from among possibly several alternatives when spreading white upwards under the constraint
White Lock.
This chapter specifies a contraction algorithm in three successively less greedy
versions, where greediness (or efficiency) is sacrificed in the interests of ensuring minimal mutilation of the agents belief system. We prove that the least
greedy version of the algorithm is minimally mutilating in a suitable sense.
142
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
Given a network T, and a node p within it, one could set about finding a
contraction in the following unintelligent, albeit completely mechanical, way.
Remember that T is a finite set of steps, and that any contraction of T with
respect to p is a subset of T. So, to find a contraction in a really plodding fashion,
proceed as follows.
Consider in turn each proper subset R of T. R must pass tests (1), (2) and (3) below in
order to count as a contraction of T with respect to p. (Failure on any one of these tests
prompts a different choice of subset of T.)
1. Check that R is a T-closed subnetwork of T.
2. Check that |R| does not contain p.
3. Consider in turn each non-empty set S of steps such that R S T. Apply tests
(1) and (2) to S. If some such S passes tests (1) and (2), then R fails test (3). But
if every such S fails either test (1) or (2), then R passes test (3).
It may take a month of Sundays, but this algorithm will eventually turn up a
contraction of T with respect to p. In similar vein one would eventually find
ones keys in a car park by beginning under a lamppost nowhere near ones past
route, and searching meticulously in an ever-widening spiral.
What we clearly want, by contrast, is a more discerning algorithm: one that
goes to work on p, and pays immediate attention to ps place within the dependency network. This is like searching for ones lost keys by beginning at ones
parked car, and re-tracing ones earlier steps away from and back to the car as
best one can.
A B RU T E - F O R C E , U N D I S C E R N I N G A L G O R I T H M
143
5.3 Preliminaries
We shall now introduce some further definitions that will be of use in stating
a general (non-deterministic) algorithm for the contraction process. We are
dealing with dependency networks T, R, etc., involving nodes x, y, etc.
In Section .. we specified an algorithm for what we called the universal
closure of a belief scheme B within a set C of steps and with respect to a set P
of nodes not in B. We said that this algorithm determines the universal closure
because of the word each in the while condition. The algorithm is used in order
to expand the belief scheme B upon adopting as new beliefs all the members
of P. This expansion can of course draw upon all the (transitional) steps known
to the agent, which make up the set C.
We now wish to specify a related algorithm, this time for identifying the set
of all steps in B that would have to be whitened if one were told that certain
nodes within B had been whitened. Let Q be the set of these initially whitened
nodes in B.
Algorithm 5.3.1: E x i s t e n t i a l C lo s u r e ( B, Q )
Steps B
Closure the set of steps in B with a premise in Q
WhiteNodes Q
while the conclusion of some step in Closure
is not in WhiteNodes
and is not the conclusion of any step in Steps
WhiteNodes
WhiteNodes NewWhiteNodes
Closure
Closure
NewSteps
We say that this algorithm determines the existential closure of B with respect
to Q because of the phrase at least one in the while condition. Existential
closure is the operation we use in order to determine which steps to whiten
in a Downward pass of our contraction algorithm.
144
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
The reader should note carefully the direction of the arrows in the top and
middle part of the diagram. The single node c below the top two has two arrows
coming into it, one vertically downwards from the (inference stroke below the)
two nodes immediately above it, and one vertically upwards from the (inference
stroke above the) two nodes immediately below it. Bear in mind that when we
speak of a downward sweep or pass (of whitening), we intend downward to
be understood as in the direction in which arrows point.
The following sequence of increasingly whitened diagrams (to be read leftto-right and top-down) exhibits the stepwise application of the algorithm for
determining the existential closure of T with respect to the set X consisting of
the two initially whitened white nodes a and b. The termination of the algorithm represents a completed Downward pass, in the sense of downward just
explained.
When an inference stroke is whitened that represents placing the step in
question into Closure. (Closure will always contain only white steps, i.e. steps
whose inference strokes are white.) The set Steps consists of those steps whose
inference strokes remain black. At each stage the membership of the set WhiteNodes is obvious. At each stage at which the while condition applies, it will be
equally obvious what nodes find their way into the set NewWhiteNodes. Each
succeeding diagram in the sequence represents the new stage reached upon
performing the do clause in response to ascertaining that the while condition
is satisfied.
PRELIMINARIES
145
Above left: The two newly whitened nodes (in the second tier) have been
whitened because they stand as conclusions of only white steps. Note, however,
that the node c is not whitened. This is because c stands as the conclusion of a
black step, i.e. the one whose inference stroke is above it.
146
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
Above right: We now whiten the two lowest inference strokes, since each one
has a premise set containing at least one white node.
Finally (as shown in this last diagram) we whiten the two bottommost nodes,
since they stand as conclusions of only white steps.
At this stage, the while condition of our algorithm, namely:
the conclusion of some step in Closure is not in White Nodes and is not the conclusion of
any step in Steps
that is,
the conclusion of some white step is black and is not the conclusion of any black step,
or, equivalently,
some black node is the conclusion of only white steps
147
T (X) will contain all nodes whose presence (as conclusions) in T can be justified only by appeal, ultimately, to the presence in T of some member of X.
Recall Definition .. in Section .. of the kernel |T| of a dependency
network T: it is the set of nodes x such that for some U, T contains the step
(U, x).
Definition 14 x is a T-casualty for R if and only if x is in |T| but not in |R|.
Definition 15 A cluster is a non-empty finite family of non-empty finite sets of
nodes, none of which properly includes any other.
Here is an example of a cluster (in which the six large rectangles represent
members of the cluster, and the circles within them represent their members):
148
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
PRELIMINARIES
149
150
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
whitens them (i.e. whitens their inference strokes). For the T-steps represent
inferential transitions, which hold because of logical, mathematical, conceptual
or evidential relations and connections that the agent has discovered thus far
among sentences and their constituents. On the present, consciously limited
account, the T-steps are being treated as though they are a priori for the agent.
Once discovered, they are never given up, in the sense of being epistemically
abandoned, or taken to be confuted. Rather, under the self-imposed restrictions
of the present account, it is only individual beliefs (i.e. nodes) that can be given
up during a contraction.
The steps in which those nodes are involved as premises and/or conclusions,
however, remain. They are never given up. Thus whenever we talk of (the rational agent, or ourselves on her behalf) getting rid of a step, or whitening it, what
we really mean is that that step no longer consists of premises that are all black
(believed by the agent) and a conclusion that is, therefore, also black (believed
by the agent). At the very least, one of its premises will be white (not believed by
the agent). This means that, for the agent, the step cannot be used to propagate
belief (from all its premises) to its conclusion. Only when all the premises are
once again believed by the agent will the step be engaged so as to support its
conclusion (for the agent). Indeed, the step will then not only be able to be thus
engagedit will have to be engaged. This is a demand of rationality that we
impose upon the agent: she is beholden to all the steps that she has recognized
as transmitting support from their premises to their conclusions. Remember:
she is a computational saint, even if not a logical one.
Note further: during a contraction process, steps are only ever given up in
this sensetaken out of the dependency network that represents the agents
former beliefson a Downward pass of the contraction algorithm. Moreover,
nodes are only ever whitenedmeaning that they are no longer believed by the
agenton an Upward pass of the contraction algorithm.
Example of a contraction: Take one of the simple examples of a dependency
network given earlier: {({a}, a), ({b}, b), ({a, b}, c)}. Call this T. One contraction
of T with respect to c is the subnetwork {({b}, b)}; another is {({a}, a)}. Note that
the contraction cannot be simply {({a}, a), ({b}, b)}, the result of whitening all
T-steps of the form (U, c) from T, as though pretending that they had never
been discovered; for the latter subnetwork is not T-closed. By reference to
T this subnetwork {({a}, a), ({b}, b)} could automatically be T-closed so as to
contain ({a, b}, c) (thereby becoming, in fact, identical to T), contrary to the
basic requirement on a contraction. Note that the empty network does not count
PRELIMINARIES
151
152
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
sets like U. We use the term excision set when it would be cumbersome to
specify the cluster in question.) We speak here of Upward whitening because of
the way in which, in our colorful method of graphical representation, premises
(in the set U = {u , . . . un }) are usually depicted as being above the conclusion
p that depends on them:
. . . un
u1
p
Whiten node p . . .
u1
...
. . .un
p
. . . hence whiten
some premise node
u1
...
. . .un
p
. . . hence whiten
the inference stroke
153
no black step with a whitened conclusion. Hence the output (of surviving black
steps) will involve no white.
We turn now to discuss in more detail the choice of an excision set, a matter
that we foreshadowed above. Note that by the formal definition of a dependency
network in Section .., no p-implying premise set can be properly included
in any other. If {U , . . . , Un } is the family (or cluster) of all the p-implying sets
( i n), then none of these sets is properly included in any other.
If we have occasion, on an Upward pass, to consider {U , . . . , Un } for collective disabling, our task will be to find an excision set (i.e. an occluder) for
the cluster {U , . . . , Un }, and whiten all of its members. The algorithm then
proceeds to work through the consequences of ones having thus whitened
these members. This it will do on its next Downward pass, which will involve
whitening (the inference stroke of) any step that has a whitened premise.
We shall presently describe an algorithm for contracting a dependency network T with respect to one of its nodes p. The contraction will proceed in stages.
As the reader will see from the description of our algorithm, we could take the
initial problem to be that of contracting a dependency network T with respect
to any (finite) set X of its nodes. But we simply follow the usual custom here of
contracting, at the outset, with respect to only one node p.
At any stage there will be four components, dynamically evolving during the
computation. These components will be called R, M, M and E.
1. The component R starts off as the given network T to be contracted, and at
the completion of each successive Downward pass R will be an ever smaller
subnetwork of T. It is not until the final value of R is attained, however, that R
becomes T-closed. The final value of R, when the computation terminates, will be
the contraction Tp as determined by that particular course of computation.
(Since the computation is non-deterministic, we have to bear in mind that Tp
is not always uniquely determined.) We choose the letter R because it reminds
us that we are dealing here with the residue of the dependency network T as
the contraction proceeds.
2. The component M is the set of nodes most recently whitened at the given
stage of computation.
3. The component M# consists of nodes whitened at the most recent stage but
one. The whitened nodes (in M or in M# ) are about to be thrown out, but
have not been just yet.
4. The component E denotes the set of nodes of T that actually have been
whitened by that stage of the contraction process, and that therefore no longer
154
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
feature as nodes of the residual network R; that is, they are not in |R|; they have
been eliminated.
Every node that gets whitened and placed in M at any Upward pass winds up
being transferred to E by the next Downward pass but one, possibly via a sojourn
in M ; but not every node in E need have been whitened beforehand in order
to wind up in E. The significance of this fact will emerge in due course.
Remember that a dependency network (such as T or any of its subnetworks R) is just a set of steps of the form (U, x) where x and all members of the
finite set U are nodes. There is no explicit representation of the inference stroke
of a step. Inference strokes are just visual aids that appear in the colored diagrams
that represent dependency networks. From what we have just said, the motivation
for the following definition should now be clear. (R, M, M , E) is to be the state
of play at any stage in the process of contracting T with respect to a node.
Definition 19 Let T be a dependency network, with R a subnetwork of T; and let
M, M and E be sets of nodes. We say that (R, M, M , E) is T-orderly if and only
if |R| and E partition |T| and M and M are disjoint subsets of |R|:
|R |
M
M#
|T |
C O N T R AC T I O N A L G O R I T H M : F I R S T V E R S I O N
155
156
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
Note how we have transferred all the pi from the M slot to the M
slot. They are now old whitened nodes, having been displaced by the
whitened newcomers in Y. Note also that the very next step of Downward will whiten all members of M .
End of Algorithm
In this algorithm, the subproblem find an occluder for X is NP-complete.
Hence the algorithm we have given is at least NP-hard. We exploited this observation in our proof of Theorem in Section ..
|a
|b
a, b | c
That is, T consists of two axioms a and b, and has developed just one consequence, namely c, from {a, b}. Consider the problem of contracting T with
respect to c.
|a
|b
a, b | c
C O N T R AC T I O N A L G O R I T H M : F I R S T V E R S I O N
157
|R|
M#
M
c
|T |
b
a
M#
M
a
c
|T |
and pass it to Downward. (Note that we have not yet actually eliminated anything, that is, put anything into the E-term. And this is the first round of
whitening, so M is still empty.)
On this new pass of Downward, things finally happen. We now have a member of the component M = {a, c} that features as a premise in R: for a is in
the implier (i.e. the premise set) {a, b} of the R-step ({a, b}, c); and it is also
in the implier of the R-step ({a}, a). Accordingly by Downward(II), these steps
are deleted from R (that is, whitened within R). The resulting subnetwork is
R = {({b}, b)}. Neither the conclusion a nor the conclusion c of the whitened
steps occurs as the conclusion of any step in R . Thus both a and c are deleted
158
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
from the M-term {a, c} to leave as M ; and they are added to the E-term to
yield {a, c} as E :
({({b}, b)} , , , {a, c})
| R1 |
b
a
M1
E1
M#
a
c
|T |
C O N T R AC T I O N A L G O R I T H M : F I R S T V E R S I O N
159
Upward
pass
whitens b
Downward
pass
whitens b|c
Upward
pass
whitens a
Downward a
pass
whitens
a|a, b|c
By delaying the whitening of inference strokeswhich is bound to happen, anyway, since at least one premise node above each such stroke has been
whitenedwe underscore the fact that the main data type on which we are
focussed is the step rather than the node. In our formal modeling, there is
no atomic constituent corresponding to an inference stroke. The step from a
and b to c, say, is formalized as ({a, b}, c), or, in the corresponding list notation
of Prolog, as [[a, b], c]. Thus the inference stroke in our colored diagrammatic
representation of this step:
a
really only serves to emphasize the second comma, which separates the premise
set from the conclusion. It is appropriate, therefore, to whiten the inference
stroke only as a sign that the whole step has been removed.
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
|a
a|b
|c
a|d
b, c|d
C O N T R AC T I O N A L G O R I T H M : F I R S T V E R S I O N
161
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
That is, the Downward pass eliminates everything, leaving us nothing. It cuts
the network back to the empty set.
A moments reflection shows, however, that an effective contraction need not
be that mutilating. There is no reason at all why c should not survive while a,
b and d are eliminated. The trouble, however, is that the greedy algorithm is
unable to see this, intent as it is on identifying the d-implying premise sets as {a}
and {b, c} and then plumping (albeit non-deterministically) in an uninformed
manner for {a, c} as the excision set. For look what could have happened had
the algorithm tarried awhile to consider the possibility of whitening, not the
whole of {a, c} (and then proceeding to the Downward pass) but rather an
appropriately chosen non-empty proper subset of {a, c}, and then looking ahead
at the projected result of the next Downward pass. Had {a} been thus chosen
instead of {a, c}:
C O N T R AC T I O N A L G O R I T H M : F I R S T V E R S I O N
163
then it would have been discovered at the very next Downward pass that b would
be whitened (since b depends essentially on a):
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
pass is ineluctable; and we should be aiming to minimize that effect, not aiming
to minimize the excision set that triggers that effect.
The situation with the greedy algorithm for theory contraction can be summarized as follows. Often we shall include unnecessary members in an excision
set during the Upward pass: members (like c in our example) that seem perfectly
well chosen, and therefore perfectly eligible, so to speak, locally, for whitening;
but that, even if they were not whitened at this Upward pass, would be losing
anyway some of their fellow members in premise sets at the very next Downward pass, and could therefore escape whitening without any net loss in the
(minimally mutilating) contraction effect sought.
This general possibility prompts one to consider more carefully, at an Upward
pass, exactly which excision set is to be chosen. We do not want to choose too
large an excision set Y (that is, an occluder for the set of impliers at that stage).
For, as we have seen, we might in our haste be including in such an excision set Y
certain unnecessary members cmembers c even in whose absence from the
excision set an adequate hitter will in effect be completed by the next Downward
pass triggered by a (non-empty) proper subset Z of Y not containing c. Such a
situation has the general features
(i) Y occludes X;
(ii) Z is a proper subset of Y, hence does not hit X; but
X)) hits X.
(iii) Z ((Z) (
X))
165
Definition 21 CT (x, y, Z), the x-baffle for y in T based on Z, is the set of those
sets U such that x is the sole member of Z in U and (U, y) is in T. The union of
such a baffle contains no T-allies of x for y from Z.3
We shall use the variable B for sets of baffles.4 We shall often suppress mention
of T when T is clear from the context. By Remark , note that an x-baffle for y
(in T) need not be a singleton.
166
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
where further criteria need to be applied in order to guarantee uniqueness of the choice.) Pass the T-orderly (Q, Z, {p , . . . , pn } M , E) to
Downward.
End of Algorithm
|a |b |c |d |e
a, b | h c, d | h e, f | i
g, h | j d, i | j
|f
|g
167
downward amplification of {i} fails to intersect {g, h}. Finally, the downward
amplification of {d} fails to intersect {g, h}, since, although the step ({c, d}, h)
would be lost in the Downward pass, the step ({a, b}, h) would remain, ensuring
that h survived. We therefore have to resort to one or other of the full excision
sets. So suppose we (non-deterministically) choose {h, i}:
That ends the first Upward step. We have chosen h and i for whitening. j is now
old whitened (hence rendered as gray), and transferred from the M slot to the
M slot.
The next Downward pass whitens the steps
({g, h}, j)({d, i}, j)
and places j in E, since j has now lost all its support. (It is a conclusion only of
white steps.) The network has accordingly been reduced to
|a |b |c |d |e |f
a, b| h c,d | h e, f| i
168
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
|g
Note that the reduced network at this stage consists of the steps whose inference strokes are single black lines in the diagram. Asterisks are placed next to
the labels of nodes (here, h and i) that have been whitened, but whose whitening
has not yet resulted in the loss of the steps in which they appear as conclusions.
Once again we perform an Upward pass. The set X of h-impliers and/or
i-impliers is
{{a, b}, {c, d}, {e, f }}.
Suppose we choose {b, d, e} as an excision set for X. (There are of course alternatives to this choice.) Both h and i are now old white, and are transferred from
the M-slot to the M -slot. Since each member of {b, d, e} is self-supporting, the
next Downward pass reduces the network to
h
|a
|c
|f
|g
The old whitened nodes h and i are transferred to E along with the most recently
whitened b, d and e, because all these nodes now stand as conclusions only of
white steps. There are no longer any nodes left whitened in the residual network,
so the algorithm halts and this is its output.
169
h
|a
|c
e, f | i
|e
|f
|g
We will not be able to recover j, the original node with respect to which the contraction process began. To be sure, if we restore d instead of e to the contracted
network, j will be recoverable:
|a
|c
c, d | h
g, h | j
|d
|f
|g
But this is weak succour indeed. What we would ideally like, as a guarantee of
minimal mutilation, is a theorem to the effect that restoring any node lost via
whitening during an Upward pass of the contraction process will precipitate
back the original node with respect to which the contraction began (here, j).
There is indeed a way to ensure this result about minimal mutilation (our
Theorem below), but first we need to introduce some further refinement in
the whitening undertaken in an Upward pass. Our algorithm in its present form
170
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
Note that for every x in Z, this x-baffle will be non-empty, because of the way
excision sets are determined (as subsets of occluders of the cluster of impliers
of members of M).
So much for the recording of the necessary information in the form of baffles. What about its recall and use at subsequent Upward passes? Imagine now
that one is performing an Upward pass further along in the execution of the
contraction algorithm, say at stage k. Suppose at the earlier stage i (< k) the set
of newly whitened nodes was Mi . If z is in Mi then we shall, at the ith Upward
pass, have recorded the z-baffle
CT (z, y, Mi ).
yM i
At every following Downward pass, each such baffle is replaced by its surviving
subsetthe set of those of its members U for which no member of U other
than z has been whitened. The surviving subset will always be non-empty
because of the following constraint, furnished by the set B of all such up-dated
baffles, on any Upward pass:
() Ensure that the excision set Z chosen is such that for no x is the x-baffle in B hit by
(Z T (Z))\{x}.
C O N T R AC T I O N A L G O R I T H M : S E C O N D V E R S I O N
171
We see now why each baffle will, after the next Downward pass, have a nonempty surviving subset (in the sense just explained) to take its place in B.
() is an optimizing condition, and we have as yet no guarantee that it can
in general be fulfilled. An excision set meeting the condition () will be called
unbaffled. If Y is baffled, however, then its bafflement will be
{x-baffle B in B | (Y T (Y)) \ {x} hits B}.
Clearly we want bafflements to be minimized, if and whenever they are forced
upon us. Note that as proper up-dating of baffles proceeds, () becomes more
exigent on the choice of Y, since the kind of proscribed hitting is harder to avoid
on a proper subset of a baffle than on the baffle itself.
Non-empty bafflement of Y should trigger backtracking in the algorithm
until some suitable unbaffled alternative to Y is found, subject of course to all
the other earlier requirements that we have laid down. If no unbaffled candidate
can be found, then one may have to consider satisficing in this regarda topic
we shall not, however, address on this occasion.
Note that choosing an unbaffled excision set is merely extending the spirit of
the earlier condition concerning downward amplification.
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
C O N T R AC T I O N A L G O R I T H M : T H I R D V E R S I O N
173
|a |b |c |d |e
a, b | h c, d | h e, f | i
g, h | j d, i | j
|f
|g
174
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
j-impliers, which is {{g, h},{d, i}}. Once again we choose h and i for whitening
(bearing in mind, of course, that there are alternatives to this non-deterministic
choice of excision set). The set of baffles is empty. The next Downward pass
whitens the steps
({g, h}, j) ({d, i}, j)
and places j in E, since j has now lost all its support. (It is a conclusion only of
white steps.):
|a |b |c |d |e |f
a, b | h c, d | h e, f | i
|g
At this point we record the h-baffle CT (h, j, {h, i}) and the i-baffle
CT (i, j, {h, i}), which are, respectively,
{{g, h}} and {{d, i}}.
(In general, such baffles need not be singletons.)
Once again we perform an Upward pass. The set X of h-impliers and/or iimpliers is
{{a, b},{c, d},{e, f }}.
The choice of an excision set Y for X is now not so straightforward. Y will have
to contain either a or b, and contain either c or d, and contain either e or f . By
constraint (), Y cannot contain d, since {d} occludes the i-baffle {{d, i}}, and
does not contain i. So suppose Y is {b, c, e}. Both h and i are now transferred
from M to M . Thus M is now {b, c, e} and M is {h, i}. Since each member of
M is self-supporting, the next Downward pass reduces the network to
C O N T R AC T I O N A L G O R I T H M : T H I R D V E R S I O N
175
|a
|d
|f
|g
There are no longer any white nodes standing as conclusions of black steps
(the set M M is empty) so the algorithm halts and this is its output.
Now the earlier untoward result concerning contraction of this network is
averted. If we restore as a self-justifier any of the nodes that were whitened on
an Upward pass, the universal closure of the contraction with respect to the
restored node will contain j.
|a |b |d |e |f
a, b | h d | h e, f | i
g, h | j d, i | j
176
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
|g
|a |b |d |e |f
a, b | h d | h e, f | i
g, h | j d, i | j
|g
|a |b |d
a, b | h d | h
|e |f
e, f | i
|g
177
|a |b |d |e |f
a, b | h d | h e, f | i
g, h | j d, i | j
|g
|a |b |e |f
a, b | h e, f | i
|g
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
|a |e
e, f | i
|f
|g
which will be the output, since now no conclusion nodes of black steps are
whitened.
Note now how this contraction also satisfies condition () for minimal mutilation. The nodes whitened in Upward passes in the course of the contraction
were b, d, h and j. Restoring any one of them precipitates j back again. We need
only check this for b, d and h taken singly. If we restore just b we obtain
|a |b
|e
|f
|g
a, b|h e, f |i
g, h|j.
If we restore just d we obtain
|a |d
|e
|f
|g
e, f |i d|h
g, h|j d, i|j.
C O N T R AC T I O N A L G O R I T H M : T H I R D V E R S I O N
179
|f
e, f |i
|g
|h
g, h|j.
We shall prove below that this holds quite generally for the algorithm equipped
with the foresight of downward amplification of minimal occluders plus hindsight concerning which potential culprits escaped whitening. Constraint ()
provides just the right measure of discrimination needed. We shall show that
whenever we do obtain an output from this algorithm, that output is minimally
mutilated.
Theorem 5 Suppose the contraction algorithm can be executed with unbaffled
excision sets at every Upward pass in the process of contracting T with respect
to p. Let q be any node that was whitened in the process of contraction. Then p is
in [(Tp) {| q}]T .
Proof. Suppose the contraction process terminated after k stages. There would
have been k excision sets determined in the course of the contraction, one for
each stage. Consider the stage at which q was whitened. (If q = p, the result is
obvious.) At this stage q (= q ) found its way into an excision set M (assuming
now that q = p). It did so because at least one node q , say, had been whitened at
the previous stage, that is, placed in an excision set M , and q was in a surviving
q -implier U . (Note that the order relation of the numerical subscripts on
these excision sets reverses the order in which the excision sets would have
been determined!) The node q had, in turn, been whitened at the immediately
preceding stage because at least one node q , say, had been whitened, that is,
placed in the excision set M , and q was in a surviving q -implier U . These
considerations generate a sequence of sets U , U , , Um and a sequence of
nodes q = q , q , . . . , qm , qm = p such that for each i:
(1) qi is in Ui Mi (0 i < m k); and
(2) (qi+1 , Ui ) is in T (0 i m k).
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
181
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
Definition 31 Assume Y hits X. Then < (X, Y) =df {x, y | x < y and x and y
are in some member of X; and y, but not x, is in Y}.
< (X, Y) is the <-violation involved when X is some cluster of premise sets each
of which has to be disabled, and Y is a proposed excision set containing at least
one member from each premise set in X.
M A K I N G U S E O F E N T R E N C H M E N T I N F O R M AT I O N
183
X))
that are minimal among such hitters under set inclusion. Among these, choose those
whose <-violations are minimal under set inclusion.
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
185
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
5.10 Conclusion
We have specified in this chapter various successively less greedy versions
of a contraction algorithm. In Chapter we provide a Prolog program that
CONCLUSION
187
implements the first of them. One can demonstrate how such programs produce
intuitively satisfying results on all extant contraction problems canvassed in the
literature. Further experiences with applications of such programsespecially
for the less-greedy versions of the contraction algorithmto larger examples
can be expected to yield more material for theoretical reflection.
188
S P E C I F I C AT I O N O F A C O N T R AC T I O N A L G O R I T H M
CHAPTER 6
A Prolog Program
for Contraction
189
The queries are read, echoed, the predicate contract/ called on to answer them
(see below) and the result displayed.
These two clauses re-define certain system-defined predicates:
?- op(,fx,@).
?- op(,xfx,?-).
The reader is advised that in this presentation we eschew altogether Prologs
new-line command nl. Its many occurrences take up too much space, and are
distracting. In the actual program that we run, the command is used frequently,
in order to achieve readable output on the screen.
Here is the clause for the predicate @:
@(FileName) :see(FileName),
repeat,
read(Term),
(if isEof(Term)
then seen
else Term = (Steps ?- WhitenedSentences),
contract(Steps,WhitenedSentences,Residue),
fail).
This is the clause for the three-place start-up predicate contract:
contract(X,A,D) :listToSet(X,Y),
listToSet(A,B),
writeproblem(Y,B),
now contract(Y,B,D),
print(ANSWER:),
printsteps(D),
write(),
fail.
Use this line for multiple answers to each problem.
!.
Use this line for the first answer to each problem.
?- lib osets.
190
A P R O L O G P R O G R A M F O R CONTRACTION
T H E TO P - L E V E L PA RT O F T H E P R O G R A M
191
A P R O L O G P R O G R A M F O R CONTRACTION
printsetsofpremises(Cluster),
occluder(Cluster,Occluder),
print(for which I choose the following occluder: ),
printsentences(Occluder),
print(and then change to Downward mode ),
contraction(down,System,Occluder,Contraction).
T H E TO P - L E V E L PA RT O F T H E P R O G R A M
193
194
A P R O L O G P R O G R A M F O R CONTRACTION
passesand(R,X,T,Unwanted) :union(R,X,S),
pc(X,S),
!,
cdm(T,S),
!,
free of(Unwanted,S),
!.
passesand(R,X,T,Unwanted) :!,
fail.
Note: pc(X,S) means that every premise of a step in X is the conclusion
of some step in S. This suffices as a test for Ss being a subnetwork of T,
since we already know that R is one.
pc([ ], ) :- !.
pc([H|T],[ ]) :- !,
fail.
pc([H|T],Y) :pc this step(H,Y),
!,
pc(T,Y),
!.
pc this step([Prem,Con],Y) :pc this node(Prem,Y),
!.
pc this node([ ], ) :!.
pc this node([H|T],Y) :conc(H,Y),
!,
T H E TO P - L E V E L PA RT O F T H E P R O G R A M
195
pc this node(T,Y),
!.
conc(H,Y): H is the conclusion of some step in Y.
conc(Conclusion,[ ]) :!,
fail.
conc(Conclusion,[[ ,Conclusion]| ]) :!.
conc(Conclusion,[H|T]) :!,
conc(Conclusion,T).
cdm(T,R): R is closed downwards modulo T; that is, for any step [U,X]
in T, if every node in U is a conclusion of a step in R, then [U,X] is in R.
cdm([ ], ).
cdm([[U,X]|Tail of T],R) :member([U,X],R),
!,
cdm(Tail of T,R).
cdm([[U,X]|Tail of T],R) :black in(R,U),
!,
fail.
cdm([[U,X]|Tail of T],R) :!,
cdm(Tail of T,R).
black in(R,U): every node in U is the conclusion of some step in R.
196
A P R O L O G P R O G R A M F O R CONTRACTION
T H E TO P - L E V E L PA RT O F T H E P R O G R A M
197
setsFrom([H|T],thatHave(X,[H|SetOfTailHits],andthe(SetOfTailMisses))) :member(X,H),
!,
setsFrom(T,thatHave(X,SetOfTailHits,andthe(SetOfTailMisses))).
setsFrom([H|T],thatHave(X,SetOfTailHits,andthe([H|SetOfTailMisses]))) :setsFrom(T,thatHave(X,SetOfTailHits,andthe(SetOfTailMisses))).
intersection([H],H).
intersection([H|[K|T]],Intersect) :intersection([K|T],IntersectKT),
intersect(H,IntersectKT,Intersect).
occludes(Cluster,Occluder) :!,
hits(Occluder,Cluster),
hitsperfectly(Occluder,Cluster).
hitsperfectly(Occluder,Cluster) :member(X,Occluder),
delElement(X,Occluder,WouldBeOccluder),
hits(WouldBeOccluder,Cluster),
!,
fail.
hitsperfectly(Occluder,Cluster).
remainingwhite(System,[ ],[ ]).
remainingwhite(System,[H|T],[H|U]) :conc(H,System),
!,
remainingwhite(System,T,U).
remainingwhite(System,[H|T],U) :!,
remainingwhite(System,T,U).
198
A P R O L O G P R O G R A M F O R CONTRACTION
not member(Conclusion,Whitened),
not conc(Conclusion,System),
!,
freshlywhite(Whitened,System,T,U).
T H E TO P - L E V E L PA RT O F T H E P R O G R A M
199
freshlywhite(Whitened,System,[H|T],U) :!,
freshlywhite(Whitened,System,T,U).
infected(Whitened,[ ],[ ]).
infected(Whitened,[[Premises,Conclusion]|T],[[Premises,Conclusion]|U]) :listToSet(Whitened,Whitened),
listToSet(Premises, Premises ),
intersect(Premises,Whitened),
!,
infected(Whitened,T,U).
infected(Whitened,[H|T],U) :!,
infected(Whitened,T,U).
prem([ ], ) :!,
fail.
prem( ,[ ]) :!,
fail.
prem(Whitened,[[Premises,Conclusion]|T]) :listToSet(Whitened,Whitened),
listToSet(Premises, Premises ),
intersect(Premises,Whitened),
!.
prem(Whitened,[[Premises,Conclusion]|T]) :prem(Whitened,T).
all premise sets([ ], ,[ ]).
all premise sets( ,[ ],[ ]).
200
A P R O L O G P R O G R A M F O R CONTRACTION
T H E TO P - L E V E L PA RT O F T H E P R O G R A M
201
202
A P R O L O G P R O G R A M F O R CONTRACTION
6.3 Pre-processing
The foregoing program will work just fine, provided, of course, that the input
problems are well formed. This means, in particular, that the list to the left of
each question mark of a posed problem is a list of steps that really do form a
dependency network, strictly defined. If there is any worry that this might not
be the case, one can use the clauses below to pre-process the inputs, in order
to ensure that they are properly posed contraction problems.
PRE-PROCESSING
203
204
A P R O L O G P R O G R A M F O R CONTRACTION
PRE-PROCESSING
205
206
A P R O L O G P R O G R A M F O R CONTRACTION
printpremises(H),
print(]),
!.
printsetsofpremises([H|T]) :print([),
printpremises(H),
print(]),
print(, ),
printsetsofpremises(T).
printpremises([ ]).
printpremises([H]) :display(H),
!.
printpremises([H|T]) :display(H),
print(, ),
printpremises(T).
C L A U S E S F O R T H E VA R I O U S P R I N T C O M M A N D S
207
CHAPTER 7
The reader will see that some of the networks are repeated, using different
choices of whitened node. The subscripts in some of the network names are
intended to remind the reader where, within the network concerned, the initially whitened node is situated. (Of course, the initially whitened node need
not be an initial node!)
In all the networks we indicate initial nodes by means of arrows coming to
them from (black) inference strokes that in turn receive no arrows. One of the
networks, though it looks like a square, is called Circle because it is representative of coherentist circles of mutually supporting nodes. Any plurality of nodes
would do; we have chosen here to deal with four.
Some of the networks will strike the reader as extremely trivial. But such
simple or degenerate examples are important aids to ascertaining whether ones
algorithm, and its implementation in ones chosen programming language,
work correctly. Every programmer knows how important it is to use lists of
208
R E S U LT S O F RU N N I N G O U R P R O G R A M
Solo
The challenge with Solo is to ensure that ones program produces the empty
network as the desired contraction.
S O M E S I M P L E C O N T R AC T I O N P R O B L E M S
209
Duo
The challenge with Duo is to ensure that ones program does not get rid of b in
the desired contraction.
a
Circle
The challenge with Circle is to ensure that ones program completes it, and
whitens all the nodes, yielding the empty network as the desired contraction.
c
[[[b], b], [[c], c], [[b], a], [[c], a]] ?- [a].
Veebottom
The challenge with Veebottom is to ensure that both sources of support for a are
whitened, yielding the empty network as the desired contraction.
Veetop
210
R E S U LT S O F RU N N I N G O U R P R O G R A M
The challenge with Veetop is to ensure that ones program respects the fact that
a enjoys support from c, even when b cannot be counted on for support. So the
contraction should contain both c and a.
a
[[[a], a],[[a], b],[[a], c]] ?- [a].
Wedgetopmiddle
b
The challenge with Wedgetopmiddle is once again to obtain the empty network
as the desired contractionsince a is the only source of support indicated both
for b and for c.
a
[[[a], a], [[a], b], [[a], c]] ?- [c].
Wedgebottom
Zigzagbottommiddle
c
[[[a], a], [[b], b], [[a], c], [[a], d], [[b], d], [[b], e]] ?- [d].
The challenge with Zigzagbottommiddle isas the reader will no doubt have
guessedto yield the empty network as the desired contraction.
S O M E S I M P L E C O N T R AC T I O N P R O B L E M S
211
Zigzagtop
c
[[[a], a], [[b], b], [[a], c], [[a], d], [[b], d], [[b], e]] ?- [a].
The challenge with Zigzagtop is one of damage controldo not surrender any
of b, d or e!
Zigzagbottomend
c
[[[a], a], [[b], b], [[a], c], [[a], d], [[b], d], [[b], e]] ?- [e].
Twigbottom
c
R E S U LT S O F RU N N I N G O U R P R O G R A M
Twigtop
The challenge with Twigtop is to have the whitening go only downwards, and
not upwards thereafter. The node a must survive (i.e. remain black).
Double Twigbottom
e
[[[a], a], [[b], b], [[c], c], [[d], d], [[a, b], e], [[c, d], e]] ?- [e].
The challenge with Double Twigbottom is to generate the four distinct contractions that the reader will no doubt be able to identify.
Double Twigtop
e
[[[a], a], [[b], b], [[c], c], [[d], d], [[a, b], e], [[c, d], e]] ?- [b].
The challenge with Double Twigtop is once again one of damage control. All
that is black (apart from the inference strokes above and below b) must remain
black.
S O M E S I M P L E C O N T R AC T I O N P R O B L E M S
213
Sprigbottom
[[[a], a], [[b], b], [[c], c], [[d], d], [[a, b], e], [[c, d], f ], [[e, f ], g]] ?- [g].
The challenge with Sprigbottom is to whiten one whole branch at a time, thereby
generating the four distinct -node contractions that the reader will easily identify. (Each of these contractions will have distinct black nodes.)
Sprigtop
g
[[[a], a], [[b], b], [[c], c], [[d], d], [[a, b], e], [[c, d], f ], [[e, f ], g]] ?- [b].
214
R E S U LT S O F RU N N I N G O U R P R O G R A M
Hexagonbottom
f
[[[a], a], [[b], b], [[c], c], [[a, b], d], [[b, c], e], [[d, e], f ]] ?- [f ].
Hexagontopmiddle
[[[a], a], [[b], b], [[c], c], [[a, b], d], [[b, c], e], [[d, e], f ]] ?- [b].
S O M E S I M P L E C O N T R AC T I O N P R O B L E M S
215
d
Hexagontopend
[[[a], a], [[b], b], [[c], c], [[a, b], d], [[b, c], e], [[d, e], f ]] ?- [a].
The challenge with Hexagontopend is contain the spread of white to just the
branch adf .
Rhombusbottom
d
[[[a], a], [[b], b], [[c], c], [[a, b], d], [[b, c], d]] ?- [d].
The challenge with Rhombusbottom is to ensure that d does not get back in again.
Whitening b will ensure that. So too will whitening both a and c.
Rhombustopmiddle
d
[[[a], a], [[b], b], [[c], c], [[a, b], d], [[b, c], d]] ?- [b].
216
R E S U LT S O F RU N N I N G O U R P R O G R A M
Rhombustopend
d
[[[a], a], [[b], b], [[c], c], [[a, b], d], [[b, c], d]] ?- [a].
The challenge with Rhombustopend is to refrain from whitening any other node.
Chaintop
Chainbottom
[[[a], a], [[a], b], [[b], c], [[c], d], [[d], e]] ?- [a].
[[[a], a], [[a], b], [[b], c], [[c], d], [[d], e]] ?- [e].
The challenge with each of these chains is to generate the empty network as the
desired contraction, but differently in each case. With Chaintop , all the whitening is accomplished in a Downward pass. With Chainbottom , all the whitening
S O M E S I M P L E C O N T R AC T I O N P R O B L E M S
217
problems
Solo
Script started on Wed Mar 24 22:56:04 2010
mercutio:/prooffinders [1] np
NU-Prolog 1.6.5
1?- [ctc].
Consulting /n/mercutio/1/neilt/prooffinders/ctc.nl.
Consulting /n/mercutio/1/neilt/prooffinders/
ctcmanagershort.nl.
Warning in op(1200, xfx, (?-)) -- changing a
system-defined operator.
Loading /n/mercutio/0/nuprolog/lib/lib.10605/osets.
no.
done
true.
2?- t(solo).
What is the result of contracting the system
a |- a
with respect to the set of sentences
[a] ?
Now I am in Downward mode.
Residual system:
a |- a
Whitened sentences:
a
Existential closure:
a |- a
New Residual system:
[ ]
Remaining whitened sentences:
218
R E S U LT S O F RU N N I N G O U R P R O G R A M
[ ]
Final contraction is:
[ ]
ANSWER:
[ ]
####################################################
Thank you for your attention.
With the remaining examples, the output will be shorn of all unnecessary
computerese, along with unnecessary parts of the automatically running commentary on the execution of the program. Bits of the latter will be left in only
when they help to illuminate some new feature of the example.
Duo
ANSWER:
b |- b
Circle
What
a |b |c |d |with
[a]
ANSWER:
[ ]
Veebottom
What
b |b |c |-
219
c |- a
with respect to the set of sentences
[a] ?
ANSWER:
[ ]
Veetop
What
b |b |c |c |with
[b]
ANSWER:
c |- c
c |- a
Wedgetopmiddle
What
a |a |a |with
[a]
ANSWER:
[ ]
Wedgebottom
What is the result of contracting the system
a |- a
220
R E S U LT S O F RU N N I N G O U R P R O G R A M
a |a |with
[c]
b
c
respect to the set of sentences
?
ANSWER:
[ ]
Zigzagbottommiddle
What
a |a |a |b |b |b |with
[d]
ANSWER:
[ ]
Zigzagtop
What
a |a |a |b |b |b |with
[a]
ANSWER:
b |- b
PROGRAM ON THE FOREGOING PROBLEMS
221
b |- d
b |- e
Zigzagbottomend
What
a |a |a |b |b |b |with
[e]
ANSWER:
a |- a
a |- c
a |- d
Twigbottom
With this problem, we revert to showing a fuller version of the programs output,
since it involves backtracking.
What is the result of contracting the system
a |- a
a,b |- c
b |- b
with respect to the set of sentences
[c] ?
Now I am in Downward mode.
Residual system:
a |- a
a,b |- c
b |- b
Whitened sentences:
222
R E S U LT S O F RU N N I N G O U R P R O G R A M
c
Existential closure:
[ ]
New Residual system:
a |- a
a,b |- c
b |- b
Remaining whitened sentences:
c
Now I am in Upward mode.
Residual system:
a |- a
a,b |- c
b |- b
Whitened sentences:
c
leading me to identify the following cluster of
premise-sets:
[a, b]
for which I choose the following occluder:
a
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- c
b |- b
Whitened sentences:
a
Existential closure:
a |- a
a,b |- c
New Residual system:
b |- b
Remaining whitened sentences:
[ ]
Final contraction is:
b |- b
PROGRAM ON THE FOREGOING PROBLEMS
223
ANSWER:
b |- b
####################################################
(The program now backtracks to change its choice of occluder. This is the first
time we have seen this happen.)
for which I choose the following occluder:
b
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- c
b |- b
Whitened sentences:
b
Existential closure:
a,b |- c
b |- b
New Residual system:
a |- a
Remaining whitened sentences:
[ ]
Final contraction is:
a |- a
ANSWER:
a |- a
Twigtop
What is the result of contracting the system
a |- a
a,b |- c
b |- b
with respect to the set of sentences
[b] ?
224
R E S U LT S O F RU N N I N G O U R P R O G R A M
ANSWER:
a |- a
DoubleTwigbottom
With this example we see, for the first time, four alternative solutions. Each of
the last three solutions is generated upon backtracking by trying an alternative
occluder.
What is the result of contracting the system
a |- a
a,b |- e
b |- b
c |- c
c,d |- e
d |- d
with respect to the set of sentences
[e] ?
Now I am in Downward mode.
Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- e
d |- d
Whitened sentences:
e
Existential closure:
[ ]
New Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- e
d |- d
PROGRAM ON THE FOREGOING PROBLEMS
225
R E S U LT S O F RU N N I N G O U R P R O G R A M
227
DoubleTwigtop
What is the result of contracting the system
a |- a
a,b |- e
b |- b
c |- c
c,d |- e
d |- d
with respect to the set of sentences
[b] ?
ANSWER:
a |- a
c |- c
c,d |- e
d |- d
Sprigbottom
What is the result of contracting the system
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
e,f |- g
with respect to the set of sentences
[g] ?
Now I am in Downward mode.
Residual system:
a |- a
a,b |- e
b |- b
228
R E S U LT S O F RU N N I N G O U R P R O G R A M
c |- c
c,d |- f
d |- d
e,f |- g
Whitened sentences:
g
Existential closure:
[ ]
New Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
e,f |- g
Remaining whitened sentences:
g
Now I am in Upward mode.
Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
e,f |- g
Whitened sentences:
g
leading me to identify the following cluster of
premise-sets:
[e, f]
for which I choose the following occluder:
e
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
PROGRAM ON THE FOREGOING PROBLEMS
229
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
e,f |- g
Whitened sentences:
e
Existential closure:
e,f |- g
New Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
Remaining whitened sentences:
e
Now I am in Upward mode.
Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
Whitened sentences:
e
leading me to identify the following cluster of
premise-sets:
[a, b]
for which I choose the following occluder:
a
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
230
R E S U LT S O F RU N N I N G O U R P R O G R A M
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
Whitened sentences:
a
Existential closure:
a |- a
a,b |- e
New Residual system:
b |- b
c |- c
c,d |- f
d |- d
Remaining whitened sentences:
[ ]
Final contraction is:
b |- b
c |- c
c,d |- f
d |- d
ANSWER:
b |- b
c |- c
c,d |- f
d |- d
#####################################################
for which I choose the following occluder:
b
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
PROGRAM ON THE FOREGOING PROBLEMS
231
d |- d
Whitened sentences:
b
Existential closure:
a,b |- e
b |- b
New Residual system:
a |- a
c |- c
c,d |- f
d |- d
Remaining whitened sentences:
[ ]
Final contraction is:
a |- a
c |- c
c,d |- f
d |- d
ANSWER:
a |- a
c |- c
c,d |- f
d |- d
#####################################################
for which I choose the following occluder:
f
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
e,f |- g
Whitened sentences:
f
232
R E S U LT S O F RU N N I N G O U R P R O G R A M
Existential closure:
e,f |- g
New Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
Remaining whitened sentences:
f
Now I am in Upward mode.
Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
Whitened sentences:
f
leading me to identify the following cluster of
premise-sets:
[c, d]
for which I choose the following occluder:
c
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
Whitened sentences:
c
233
Existential closure:
c |- c
c,d |- f
New Residual system:
a |- a
a,b |- e
b |- b
d |- d
Remaining whitened sentences:
[ ]
Final contraction is:
a |- a
a,b |- e
b |- b
d |- d
ANSWER:
a |- a
a,b |- e
b |- b
d |- d
#####################################################
for which I choose the following occluder:
d
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
Whitened sentences:
d
Existential closure:
c,d |- f
d |- d
234
R E S U LT S O F RU N N I N G O U R P R O G R A M
Sprigtop
What is the result of contracting the system
a |- a
a,b |- e
b |- b
c |- c
c,d |- f
d |- d
e,f |- g
with respect to the set of sentences
[b] ?
ANSWER:
a |- a
c |- c
c,d |- f
d |- d
235
Hexagonbottom
What is the result of contracting the system
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
d,e |- f
with respect to the set of sentences
[f] ?
Now I am in Downward mode.
Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
d,e |- f
Whitened sentences:
f
Existential closure:
[ ]
New Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
d,e |- f
Remaining whitened sentences:
f
Now I am in Upward mode.
Residual system:
a |- a
a,b |- d
b |- b
236
R E S U LT S O F RU N N I N G O U R P R O G R A M
b,c |- e
c |- c
d,e |- f
Whitened sentences:
f
leading me to identify the following cluster of
premise-sets:
[d, e]
for which I choose the following occluder:
d
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
d,e |- f
Whitened sentences:
d
Existential closure:
d,e |- f
New Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
Remaining whitened sentences:
d
Now I am in Upward mode.
Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
PROGRAM ON THE FOREGOING PROBLEMS
237
Whitened sentences:
d
leading me to identify the following cluster of
premise-sets:
[a, b]
for which I choose the following occluder:
a
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
Whitened sentences:
a
Existential closure:
a |- a
a,b |- d
New Residual system:
b |- b
b,c |- e
c |- c
Remaining whitened sentences:
[ ]
Final contraction is:
b |- b
b,c |- e
c |- c
ANSWER:
b |- b
b,c |- e
c |- c
#####################################################
for which I choose the following occluder:
b
and then change to Downward mode ...
238
R E S U LT S O F RU N N I N G O U R P R O G R A M
239
Existential closure:
d,e |- f
New Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
Remaining whitened sentences:
e
Now I am in Upward mode.
Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
Whitened sentences:
e
leading me to identify the following cluster of
premise-sets:
[b, c]
for which I choose the following occluder:
b
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
Whitened sentences:
b
Existential closure:
a,b |- d
b |- b
b,c |- e
240
R E S U LT S O F RU N N I N G O U R P R O G R A M
241
a |- a
a,b |- d
b |- b
Hexagontopmiddle
What is the result of contracting the system
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
d,e |- f
with respect to the set of sentences
[b] ?
ANSWER:
a |- a
c |- c
Hexagontopend
What is the result of contracting the system
a |- a
a,b |- d
b |- b
b,c |- e
c |- c
d,e |- f
with respect to the set of sentences
[a] ?
ANSWER:
b |- b
b,c |- e
c |- c
242
R E S U LT S O F RU N N I N G O U R P R O G R A M
Rhombusbottom
What is the result of contracting the system
a |- a
a,b |- d
b |- b
b,c |- d
c |- c
with respect to the set of sentences
[d] ?
ANSWER:
b |- b
####################################################
(Here now is the only backtracking alternative choice of occluderNT.)
for which I choose the following occluder:
b
and then change to Downward mode ...
ANSWER:
a |- a
c |- c
Rhombustopmiddle
What is the result of contracting the system
a |- a
a,b |- d
b |- b
b,c |- d
c |- c
with respect to the set of sentences
[b] ?
243
ANSWER:
a |- a
c |- c
Rhombustopend
What is the result of contracting the system
a |- a
a,b |- d
b |- b
b,c |- d
c |- c
with respect to the set of sentences
[a] ?
ANSWER:
b |- b
b,c |- d
c |- c
Chaintop
What
a |a |b |c |d |with
[a]
ANSWER:
[ ]
244
R E S U LT S O F RU N N I N G O U R P R O G R A M
Chainbottom
One might think that this case will be as straightforward as the last one. We can
just see that white will be spread upwards in a straight line to the node a at the
top. But the algorithm does not see that. It methodically performs alternating
upward and Downward passes. The Downward passes remove just one step at
a time. The remaining whitened node, at any stage, is the bottommost node of
the residual straight line, or branch, of the original system. On each Upward
pass, the node immediately above it is whitened. On the subsequent Downward pass, the step from the higher node to the node immediately below it is
eliminated.
What
a |a |b |c |d |with
[e]
245
e
Now I am in Upward mode.
Residual system:
a |- a
a |- b
b |- c
c |- d
d |- e
Whitened sentences:
e
leading me to identify the following cluster of
premise-sets:
[d]
for which I choose the following occluder:
d
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a |- b
b |- c
c |- d
d |- e
Whitened sentences:
d
Existential closure:
d |- e
New Residual system:
a |- a
a |- b
b |- c
c |- d
Remaining whitened sentences:
d
Now I am in Upward mode.
Residual system:
a |- a
a |- b
246
R E S U LT S O F RU N N I N G O U R P R O G R A M
b |- c
c |- d
Whitened sentences:
d
leading me to identify the following cluster of
premise-sets:
[c]
for which I choose the following occluder:
c
and then change to Downward mode ...
Now I am in Downward mode.
Residual system:
a |- a
a |- b
b |- c
c |- d
Whitened sentences:
c
Existential closure:
c |- d
New Residual system:
a |- a
a |- b
b |- c
Remaining whitened sentences:
c
Now I am in Upward mode.
Residual system:
a |- a
a |- b
b |- c
Whitened sentences:
c
leading me to identify the following cluster of
premise-sets:
[b]
for which I choose the following occluder:
b
PROGRAM ON THE FOREGOING PROBLEMS
247
R E S U LT S O F RU N N I N G O U R P R O G R A M
249
PA RT II
CHAPTER 8
253
the conclusion that there is a natural number n such that F(n), it would be
sufficient simply to assume that no natural number has the property F, and then
(classically!) derive an absurdity from that assumption:
(i)
xF(x)
..
.
(i)
xF(x)
Thus the intuitionists ended up rejecting the rule of Classical Reductio ad Absurdum (CR):
(i)
..
.
(i)
and all rules equivalent to it, modulo the set of rules that the intuitionist could
eventually motivate or justify in a more direct fashion. Among these intuitionistic equivalents of (CR) is the Law of Excluded Middle (LEM):
through whose rejection intuitionistic logic is perhaps better known.
But even the intuitionists retained the rule Ex Falso Quodlibet (EFQ), also
known as the Absurdity Rule:
which allows one to infer any conclusion one wishes as soon as one has derived
an absurdity. This residual rule within intuitionistic logic is anathema to relevantists, since it affords an easy proof of the infamous Lewiss First Paradox:
A, A B. The proof is
A
254
Relevantists refuse to accept Lewiss First Paradox, on the grounds that there
need not be any connection in meaning between the sentence A in its premises
and its conclusion B. Relevantists regard such a lack of relevance between the
premises and conclusions of certain classically (and intuitionistically) approved
rules of inference as compromising their claim to genuine validity. Many relevantists are still otherwise classical in their orientation, in endorsing (as relevantly valid) such inferences as Double-Negation Elimination (DNE), another
intuitionistic equivalent of CR (and of LEM):
Both campsintuitionist and relevantisthave variously produced philosophical, methodological, meaning-theoretic and intuitive considerations in support of their respective recommendations for restricting classical logic. Their
respective complaints about aspects of classical logic have, however, tended to
be orthogonal to one another.
Still, all participants in the debate over logical reform have an eye to the
methodological requirements of mathematics and natural science. Two central
concerns have been:
Does ones logic afford all the mathematical theorems that are needed for application in
science?
and
Does ones logic enable one to carry out the most rigorous possible tests of a scientific theory?
255
A much better name for the system in question, however, would be Core
Logic; and that is the name we shall use here.
In the following statement of the rules of Core Logic, the boxes next to discharge strokes indicate that vacuous discharge is not allowed. There must be an
assumption of the indicated form available for discharge. (With (-E) and (-E)
we require only that at least one of the indicated assumptions should have been
used, and be available for discharge.) The diamond next to the discharge stroke
in the second half of (-I) indicates that it is not required that the assumption
in question should have been used and be available for discharge. But if it is
available, then it is discharged.
a indicates that no occurrence of the parameter a is permitThe annotation
ted in the sentence(s) thus annotated.
Note, finally, that major premises for eliminations must stand proud; they are
to have no proof-work above them.
(i)
(-I)
(-E)
..
.
(i)
..
.
(-I)
..
.
..
.
(i)
(i)
,
..
.
(-E)
(i)
256
(-I)
..
.
..
.
(-E)
..
.
(i)
..
.
(i)
(-I)
(i)
..
.
(i)
..
.
(-E)
..
.
..
.
(i)
..
.
(i)
(i)
..
.
(i)
..
.
(i)
(i)
(i)
..
.
(-I)
(i)
(i)
(i)
(i)
..
.
tx
x
(i)
a. . . ax . . .
a
..
.
(-E)
a
a
x
(i)
a
(-I)
..
.
xxa
CORE LOGIC
257
(i)
tx
(-E)
. . .. . .
(i)
, . . . , txn
..
.
(=-I)
(i)
t=t
(=-E)
t=u
..
.
t
t
, where u = u and = .
, :
:
(-L)
:
, :
(-R)
:
:
, :
258
, :
, :
(-L)
, :
, :
(-R)
:
:
: :
(-L)
(-R)
(-L)
, : , : , : , : , : , :
, , :
, , :
, , :
, :
:
, :
: : :
, :
:
, , :
(-R)
: tx
: x
(-L)
, ax :
, x :
(-R)
(-L)
(=-R)
(=-L)
, , :
, :
(where )
:
where a occurs in but in no member of
: xxa
, tx , . . . , txn :
, x :
:t=t
:
, t=u :
where ut = ut
and
A,
,
259
and one would have the assurance that the resulting construction would count
as a proof of the overall conclusion from the set of combined assumptions. Finite repetitions of the accumulation operation would also of course be
countenanced:
n
n
n
A , . . . , An ,
together yield (A ) , . . . , (An ) , ,
,, n and
A
An
A
into a core proof [] of or of from (some subset of) .
260
A
An
(where A , . . . , An and may be empty) into a core proof
[ [. . . [n ] . . .]]
of or of from (some subset of) . . . n .
Thus the sequence of different kinds of justification for choice of logic will at
last be rooted in one that underscores the appropriateness of the new name for
the system. The principles of Core Logic are the core principles of logic.
261
263
conclusion is roughly of the form The way to fix this problem is to make it the
case that . . . . Let us explain.
The inference strokes b , . . . , bn send arrows to the node a, and are all the
inference strokes doing so:
Ab a, . . . , Abn a;
x((Sx Axa) (x = b . . . x = bn )).
The inference strokes b , . . . , bn are all white:
Sb , . . . , Sbn ;
Wb , . . . , Wbn .
The node a is black:
Na;
Ba.
Let us denote by the set of diagrammatic assumptions just listed. A fragmentary situation of the kind described by sins against our rules; that is the
problem on our hands. Logically, it manifests itself thus: from and certain of
our Axioms of Coloration, one can derive absurdity. That is, we can detect the
violation in question, by deducing from and appropriately chosen Axioms
of Coloration. With as stated, where the monadic predicate W is used, these
chosen axioms can be Axioms (), () and (). Alternatively, if is formulated
using B in place of W, we can get by with just Axiom (). Let us follow this
second alternative. The logical task for the revision theorist, then, is to deduce
from the set of assumptions
Bb , . . . , Bbn , Ba, Na,
Sb , . . . , Sbn , Ab a, . . . , Abn a,
x((Sx Axa) (x = b . . . x = bn )),
x((Bx Nx) y(By Sy Ayx))
This sentence is an example of one that is a logical consequence of the L-sentence that categorically describes the A-, N- and S- structure of the network (i.e. of the whole diagram), assuming
that the language L contains also the names a, b , . . . , bn .
264
Bb : Bb
b=b i , Bb : Bb i
b = b i , Bb i , Bb : in
Sb Aba : Sb Aba
ni= b = b i , Bb , . . . , Bb n , Bb :
Aba) ni=
Ba : Ba Na : Na
Ba, Na : BaNa
(Sb
b = b i , Bb, Sb Aba, Bb , . . . , Bb n :
Bb, Sb Aba, x((Sx Axa) ni= x = b i ), Bb , . . . , Bb n :
BbSb Aba, x((Sx Axa) ni= x = b i ), Bb , . . . , Bb n :
y(By Sy Aya)), x((Sx Axa) ni= x = b i ), Bb , . . . , Bb n :
Note that all steps but one within this proof are Left-introduction steps (which,
in natural deduction, would be steps of elimination). The exception is the step
of (-R) at the top on the left. This appeal to a Right-introduction rule could,
however, have been avoided, had we used the sentence
x(Bx(Nxy(BySyAyx)))
instead of
x((BxNx)y(BySyAyx)).
With that choice of a logical equivalent, we would have had a proof of inconsistency consisting only of Left-introduction steps. But even in the proof we have
displayed, every Left-introduction rule finds application. So, whatever content
is conferred on a logical operator by its Left-introduction rule, we may conclude
that it needs to be endowed with at least that much content in order for us to
be able (rigorously to) detect that, with the diagram fragment above, we have a
problem on our hands, in the form of a violation of Coloration Axiom ().
R E F L E X I V E S TA B I L I T Y
265
The inference stroke b, which is black, sends an arrow to the node a, which is
white:
Replacing, as before, W with B, we have the task of deriving from the set of
assumptions
in order to detect the violation of Axiom () that this diagram fragment represents. The reductio proof below makes do without the assumption Sb.
Bb : Bb
Bb, Bb :
Aba : Aba Bb Sb, Bb :
Ba : Ba Na : Na
Ba, Na : BaNa
b
g
?
a
Here we have a violation of Axiom (): Every black inference stroke receives
arrows (if any) only from black nodes; in logical symbols,
x((Bx Sx) y(Ayx (By Ny))).
Following the method already illustrated above, we seek a proof of from
Axiom () along with the assumptions
Nb, Bb, Aba, Sa, Ba.
The assumptions other than Nb suffice. The sought proof can be obtained from
the previous one, by substituting S for N throughout, and B for B in certain
places:
Aba : Aba
Ba : Ba Sa : Sa
Ba, Sa : BaSa
Bb : Bb
Bb, Bb :
BbSb, Bb :
@
@
R
@
a
bn
w
R E F L E X I V E S TA B I L I T Y
267
Here we have a violation of Axiom (): Every white inference stroke that receives
an arrow receives an arrow from some white node; in logical symbols,
x((Wx Sx zAzx) y(Wy Ny Ayx)).
The nodes b , . . . , bn send arrows to the inference stroke a, and are all the nodes
doing so:
Ab a, . . . , Abn a;
x((Nx Ax, a)) (x = b . . . x = bn )).
The nodes b , . . . , bn are all black:
Nb , . . . , Nbn ;
Bb , . . . , Bbn .
The inference stroke a is not black:
Sa;
Ba.
Analogously to the case with the violation of Axiom (), it turns out that we
have no need for the assumptions of the form Nbi . We do, however, need one
(but only one) assumption of the form Abi a. For definiteness, let us take Ab a.
That will guarantee zAza. The proof of from the remaining assumptions is
similar to that in the case of Axiom ():
Bb : Bb
b=b i , Bb : Bb i
b = b i , Bb i , Bb : in
Sb Aba : Sb Aba
Ba : Ba
Sa : Sa
Ba, Sa : Ba Sa
Ab a : Ab a
Ab a : zAza
ni= b = b i , Bb , . . . , Bbn , Bb :
Similar remarks as before hold for steps of Left-introduction within this proof.
268
we are prevented from doing anything that would result in our having to give
up (or deny) any of Bb , . . . , Bbn . That leaves only Ba as the possible culprit.
So it is more of a Hobsons choice, rather than an instance of the QuineDuhem
problem (which problem consists in there being more than one assumption that
could, with equal reason, be rejected). The resolution of the problem at hand
consists in holding on to all of
x((Bx Nx) y(By Sy Ayx)), x((Sx Axa) ni= x = bi ),
Bb , . . . , Bbn , Na,
R E F L E X I V E S TA B I L I T Y
269
and making it the case that we can give up Ba, hence deny it. That is, we get
ourselves into a position where we can assert Ba (which is equivalent to
asserting Wa). In order to do this, we whiten the node a. The resulting diagram
fragment is accordingly
bn
b
all white
...
HH
HH
jg
H
a
It is easy to see that this diagram fragment is in accord with the Axioms of
Coloration. That is not to say, however, that there will be no further violations
of Coloration Axioms to be resolved. For, by whitening the node a, we could
well have produced such a violation elsewhere in the diagram, at a neighboring
fragment. That indeed is how a process of contraction or expansion plays out
successive adjustments in coloration of nodes, dealing with the knock-on effects
at neighboring parts of the diagram, until, in the end (after finitely many such
steps) we achieve a coloration that is in accord with the Axioms everywhere
within the whole digram.
270
271
color of that element!, where the referent of that element (a particular node or
inference stroke) is uniquely determined.
The reasoning is practical as well as theoretical. Confronted at instant t with
a diagram fragment that is inconsistent (as shown by some core proof ) with
the set of assumptions consisting of the Axioms of Coloration and the Axioms
of Configuration, the task is to work out what to do (by way of elemental color
change), consistently with whatever Lock is in place, in order to ensure that at
the successive instant t following that elemental color change, the proof of
inconsistency cannot go through if the t -relative assumptions are substituted
for the t-relative ones. We have seen so far that with the resolutions involving
Axioms () and (), there is only one possible choice, so it is forced upon one.
The same holds for resolutions of violations of Axiom (); but, as we shall
presently see, not with violations of Axiom ().
Before passing on to examine the cases with Axiom () and (), it is worth
stressing that the resolutions under consideration are always local. They involve
diagram fragments, not the whole diagram. There is no guarantee that upon
carrying out any such local resolution there will be no further violation of
Coloration Axioms elsewhere in the diagram. Expunging all such violations
is, literally, a time-consuming task. In general, we shall pass through many
successive stages t, t , t , . . . before an equilibrium is reached in which all the
Axioms are satisfied.
Clearly if Black Lock is in effect, the only thing we can do to avoid this inconsistency is make the inference stroke a black. If White Lock is in effect, however,
we have a range of similar choices: we can avoid the inconsistency by whitening
just one of the nodes b , . . . , bn . With n possible choices of action, the process becomes non-deterministic. Each of the available choices could lead to the
process playing out in a distinctive way across the whole diagram N , thereby
resulting in many different N -wide coloration equilibria that would satisfy
not only the categorical description N of the diagram, but also the Axioms
of Coloration.
T H E U P S H OT
273
CHAPTER 9
[Boris] had always treated objects like forks and spoons as though they had been
indefinitely renewablea profound mistake; there was a finite number of bars, cinemas,
houses, towns, and villages, and one and the same individual could not go to each of them
more than a finite number of times. At the moment of his death, in 42, he would have
lunched 365 22 times, 8,030 times in all, counting his meals in infancy. And assuming
that he had had an omelet once out of every ten occasions, he would have eaten 803
omelets.Only 803 omelets?" he said to himself with astonishment. Oh no; there are
the dinners too, that makes 16,060 meals, and 1,606 omelets. Anyway, for a man partial
to omelets that was not a considerable total. And what about cafs? he continued. The
number of times that I shall go to a caf after today can be calculated: suppose I go twice
a day, and that I am called up in a year, that makes 730 times. 730 times!thats not very
often.
Sartre (1947), at pp. 2712.
s with Boriss omelets, so with our beliefs: there are only finitely many
of them. Gentle reader, do not raise your hand with your immediate objection. Of course a logical law, such as the Law of NonContradiction:
274
T H E F I N I TA RY P R E D I C A M E N T
( )
has infinitely many potential substitution instances, such as
(It is raining (It is raining)).
And of course a logical saint would be obliged to believe every one of them! Similarly, of course a simple theorem of arithmetic, such as the Law of Commutativity
of Addition:
nm n + m = m + n
has infinitely many instances, such as
+ = + .
But no logical paragonand certainly no human beinghas ever actively
entertained each and every one of these infinitely many instances. Far from it.
Even a logical paragon will have entertained only finitely many such instances,
no matter how irreproachable her thought processes have been.
We shall accordingly be interested in representations of finite stages of development of axiomatizable theories.
275
276
T H E F I N I TA RY P R E D I C A M E N T
277
T H E F I N I TA RY P R E D I C A M E N T
279
C H A P T E R 10
Mathematical
Justifications are Not
Infinitely Various
In every first-order theory axiomatized by axioms and axiom schemes, subject to mild conditions, every theorem follows logically from at most finitely many logically minimal sets of its
axioms and axiom-scheme instances.
280
M AT H E M AT I C A L J U S T I F I C AT I O N S
We leave open the question of the exact counts for PA, ZFC.
281
282
M AT H E M AT I C A L J U S T I F I C AT I O N S
10.2 Results
Theorem 7 PA has the validity property. ZFC has the validity property.
Proof. The induction axiom scheme is
(R() x(R(x) R(s(x)))) R(x).
Note that the final occurrence of x is free. Replacing R by v v = v, we obtain a
valid formula.
The separation axiom scheme is
xy(y x (R(y) y z)).
Note that the final occurrence of z is free. Replacing R by v v = v, we obtain
xy(y x (v v = v y z)),
which is valid by setting x = z. The replacement axiom scheme is
(x x y !zR(x, z))wx(x y z(R(x, z) z w)).
Note that the two occurrences of y are free. Replacing R by v v = v, and
performing some obvious logical simplifications, we obtain
(x x y !z z = z) wx(x y z z w).
To see that this is logically valid, set w = y. QED
Lemma 3 Let T be a schematic theory with the validity property, and no
axioms. Then every consequence of T is equivalent to a finite set of instances of
the schemes in T.
Proof. Let T, be as given, and let be a consequence of T. Let be a consequence of { [f ], . . . , m [fm ]}, where the s are instances of the schemes in T,
using substitution functions f , . . . , fm .
We now use substitution functions g , . . . , gm given by gi (R) = fi (R).
We will show that is equivalent to { [g ], . . . , m [gm ]}. We claim that
implies i [gi ].
To see this, we assume . Then each gj (R) holds if and only if v v = v.
Therefore each i [gi ] holds if and only if i [f ] holds, where f is constantly
v v = v. Hence each i [gi ] holds if and only if some particular logically valid
formula holds.
R E S U LT S
283
Now assume . Then each gj (R) holds if and only if fj (R). Therefore each
i [gi ] holds if and only if i [fi ] holds. Now we have that for some i, i [fi ].
Hence for some i, i [gi ] . Therefore ( [g ] . . . m [gm ]) . QED
Lemma 4 Let T be a schematic theory with the validity property, with a finite
set K of axioms. Let be a consequence of T. Then has a minimal implying set
in T.
Proof. Let T, , K, be as given. Let W be the set of all E {} such that E K
and E is a consequence of the schemes in T. Obviously K W, and W
is finite.
For each E W, we can apply Lemma . to E , to form a finite set
(E) of instances of the schemes in T that is equivalent to E . Obviously,
if E W then E (E) is equivalent to E {}. In particular, if E W then
E (E) is an implying set in T for .
Let E be a minimal element of W in the following sense. For any E W, if
E {} is a consequence of E {}, then E {} is a consequence of E {}.
We claim that E (E) is a minimal implying set in T for .
To see this, let K { ,,m } be an implying set in T for , where
K K,
m 0,
1 , . . . , m are instances of schemes in T, and
K {1 , . . . , m } is a consequence of E (E).
M AT H E M AT I C A L J U S T I F I C AT I O N S
R E S U LT S
285
PA RT III
Comparisons
C H A P T E R 11
here are two main approaches to the problem of theory contraction that
are by now reasonably well entrenched. One is from artificial intelligence; the other is from mathematical logic. The new approach that has
been adopted here is distinct from both of these, seeking to avoid different difficulties inherent in each. A section is devoted to each of these other approaches,
in order to highlight the respects in which the new approach differs.
A third section is devoted to another branch of AI, which investigates probabilistic networks.
We have chosen the foci for these sections with an eye to how closely they
address the same problems as this account, and because they are approaches
now well served by authoritative monographs on them.
The author is well aware that many alternative approaches that are represented in the journal literature could merit coverage as well. But that would lead
to a discussion of unmanageable length. By confining our attention to just these
three main approaches, we hope to put the reader in a position to extrapolate
to unexamined approaches the points of contrast and criticism that we would
be inclined to raise.
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
289
290
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
11.2 AGM-theory
The approach from mathematical logic, so-called AGM-theory, is due to
Alchourrn, Grdenfors and Makinson. (See Grdenfors [] for an account
of AGM-theory.) In this section we expound the theory, and then note some
difficulties confronting it.
AG M-T H E O RY
291
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
293
294
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
RECOVERY
11.2.3 Implementability
The first difference, on implementability, arises because we are not concerned
with (necessarily infinite) logical closures; instead, all the objects and structures with which we operate are finite, and all relations among them are effectively decidable. Moreover, there is a principled philosophical and metalogical justification for restricting our treatment to the finite and computable.
This rests on the result proved by Harvey Friedman (and presented in Chapter ), to the effect that there are at most finitely many logically minimal sets
of axioms implying any theorem, provided only that the theory in question
is axiomatized by means of finitely many axioms and finitely many normal
axiom schemes. (Every extant axiomatic theory in mathematics meets this
condition.)
Partial meet contraction functions are not, in general, computable, because
of their infinitistic character and the undecidability of the underlying relations
of logical consequence, reference to which is essential for the theoretical constructions involved. Hence AGM-theory does not offer a convincing prospect
for computer implementation of its methods of contraction.
Tennant [a] gave a preliminary and rough sketch of the staining algorithm for theory contraction. It was put forward in order to illustrate, and
make use of, a finitary conception of the structure of theories and a computational conception of the true nature of theory contractions. This computational conception was at odds with the infinitistic and non-effective conceptions
respectively presupposed and developed in AGM-theory. Grdenfors []
maintained
Although I do not develop computer implementations of the constructions, I believe that
my theory is of interest to AI (artificial intelligence) researchers. (p. ix)
these constructions are suitable for computer programs, although the implementation details are not pursued here. (p. 3)
AG M-T H E O RY
295
Moreover, AGM-theory has not found the right solution to the problem of
contraction, even in the case of theories with finite bases, in the proposed
method of base contraction. This is because (as argued in Tennant [a])
the method of base contraction is not minimally mutilating. It would appear,
then, that the promissory notes of AGM-theory just quoted have proved to be
unrealistic tender.
It is a serious shortcoming on the part of any theory of potential interest to
AI that its objects and constructions are not finitary and effective. It might be
thought that that might simply be a price that has to be paid in order to get at
the whole truth about an important range of phenomena, which may somehow
transcend the computable. If this were so in the case of theory contraction and
theory revision, perhaps AGM-theory could be forgiven for not being usefully
implementable; perhaps its real value would derive from its nevertheless being
a true picture of a very complicated matter.
Unfortunately, this is not so in the case of AGM-theory.
296
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
AG M-T H E O RY
297
298
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
such a bridge, when each end of it is blocked by a pile of rubbish? The bizarrely
irrational functions satisfy the AGM-postulates; and they can be constructed
according to the favored logico-algebraic recipe.
The situation with AGM-theory is therefore analogous to that of Alfred
J. Ayers unhappy attempt to provide a criterion of cognitive significance by
offering a definition of what Ayer called indirectly verifiable sentences (which
include the directly verifiable ones as a special case). (See Ayer [].) Alonzo
Church showed that, on a weak assumption that one would expect to be satisfied
by any scientific language, for every sentence S, either S would count as indirectly verifiable or S would count as directly verifiable, on Ayers definition.
(See Church [].) The assumption in question is that the language contains
three mutually independent observation sentences. This assumption was subsequently weakened by Ullian []. He showed that if O and O are observation
sentences such that O does not imply O, then any contingent sentence not
implied by O would count as indirectly verifiable.4
Churchs result has come to be known as a collapse proof. A collapse proof
shows that any attempted formal explication of an apparently stable and intuitively appealing informal notion (such as cognitive significance) spectacularly
misses its markeither by a surfeit of false positives or by a surfeit of false
negatives. In Ayers case, the surfeit of false positives was the whole language
every sentence counted as indirectly verifiable (on his formal definition of that
notion), while of course the target informal notion of cognitive significance was
supposed to be highly restricted in its extension.
The Degeneracy Theorem for AGM-theory shows, likewise, that its attempted
formal explication of the informal notion of a revision function seriously misses
its mark, allowing for a surfeit of false positives. Whether this deficiency can
be repaired from within the AGM-framework (by adopting more postulates)
remains to be seen. The present author is not sanguine on this score.
In Tennant [c], ch. , a theory of cognitive significance is put forward that is intended to
avoid these difficulties.
AG M-T H E O RY
299
, so that A
is [A,
]. For then
will often effect unwanted bloating in
addition to unwanted mutilation. This problem of bloating is particularly acute
when
is a complete theory, or when
has no extra-logical vocabulary (that
is, propositional atoms; or names, function signs and predicates) in common
with K or with A.
If this were not already bad enough for the AGM-theory of revision, it is
worth pointing out also that one can appeal to Theorem so as to ensure that,
for any given K inconsistent with A, and for any theory J that would definitely
count as bizarre qua revision of K with respect to A, there will be an AGMrevision function such that K A = J. The way to ensure this will be described
in Theorem .
Definition 53 Let D be a sentence and let H be a set of sentences. Then DH =df
{D | H}.
Theorem 13 Let J be a consistent theory, let D be a contingent sentence inconsistent with J, and let H be any theory that implies D. (H may be highly irrational
as a revision of J with respect to D; for H may involve both unwanted mutilation
and unwanted bloating.) For any consistent theory K and any contingent sentence
A refuted by K, set
KA =df A DH.
300
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
301
Corollary 3 Suppose that K xx. For every n > , let DKn be the claim there
are exactly n non-s. Then for each n > , one can choose a different bizarre
theory JnK (subject only to the condition that JnK be consistent with DKn ) and take
[JnK , DKn ] as the revision of K with respect to DKn .
With both these Corollaries, the bizarre extra beliefs JnK could, for example,
differ in saying, among other things, that at any given time there are exactly
n angels dancing on the head of a pin.
Corollary shows that the full set of eight AGM postulates for revision admits
of revision functions * such that for every theory K containing a universal
generalization, and for infinitely many A inconsistent with K, the revision K A
can be as bizarre as one wishes.
Note that Theorem and its two Corollaries do not presuppose, or depend
on, any particular analysis or formal explication of the notions of (minimal)
mutilation or (minimal) bloatingor of bizarreness (qua revision). The dialectical structure of the predicament in which AGM-theory is revealed to stand
is as follows. The asserter of Theorem is in effect saying You, the listener,
may choose whatever you would say is a bizarre series of revisions; then I, the
speaker, will show you a revision function that satisfies all the AGM-postulates,
yet yields all those bizarre revisions. The asserter of Theorem is allowing
his listener to work with any intuitive sense of bizarreness that she wishes. She
gets to specify exactly what would count as a bizarre K-spectrum, by her own
lights. (She might do this intuitively, without recourse to any particular logical
or set-theoretic explication of the notions of minimal bloating and minimal
mutilation.) Once she has a K-spectrum {D Ki }i< , {HiK }i< for which she
agrees that Hi would be bizarre as a revision of K with respect to Di , Theorem
can immediately be applied so as to produce a revision function * satisfying
all eight AGM postulates, but such that K Di = Hi . Moreover, as Corollary
shows, any theory K containing a universally quantified claim xx affords
one the opportunity to create such spectral mischief. One need only contemplate revisions of K with respect to the infinitely various claims to the effect
that there are, respectively, exactly , , . . . counterexamples to the universal
generalization in question.
In light of the foregoing results it would appear that an undesirable degree
of laxity has been revealed in the full set of AGM postulates for revision functions. This laxity has been revealed by attention to the postulates themselves,
rather than to the mathematical constructions afforded by the representation
theorem.
302
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
(Bear in mind that what Pearl and other probability theorists like to call variables are the propositions or events represented by the nodes.5 )
Let U be a set of propositions or events, and let P be a probability distribution
on U. Assume that where X, Y and Z are (pairwise) disjoint subsets of U, one can
assign a truth value to the probability-theoretic claim X is (probabilistically)
independent of Y given Z, where this claim of course involves reference to the
distribution P. That results in a so-called dependency model (p. ). We could
call the dependency model M (as Pearl does), and abbreviate the claim just
mentioned as I(X, Y, Z)M . But in the interests of clarity, we shall simply employ
the subscript P to the same effect, and use the abbreviation IP (X, Y, Z).
Now, if the same nodes (the ones forming the set U) were to be connected
in some pattern by arrows, so as to form a directed acyclic graphwhich we
can call Done can inquire whether the pattern of influences encoded by (the
arrows among nodes in) D accords with the given probability distribution P
on the nodesthat is, whether the claimed probabilistic (in)dependencies
IP (X, Y, Z) are respected by the pattern D of arrows. In order to diagnose
whether this is the case, Pearl introduces the following notion of separation of
pairwise disjoint subsets X, Y and Z within D (p. ). We are introducing stylistic changes here, so the following, though fair and accurate, is not a direct quote:
As Pearl has clarified (personal communication), nodes are variables, and each variable is a
partition of worlds, e.g., Variable Temperature is a partition into the events: T= T= etc.
B AY E S I A N N E T WO R K S
303
the set Z of nodes separates the set X of nodes from the set Y of nodesin short,
X|Z|YD just in case there is no path of D between any node in X and any node in Y
along which
(1) every node with converging arrows is in Z or has a descendant in Z; and
(2) every other node is outside Z.
Now suppose that for every pairwise disjoint subsets X, Y and Z of U we have
X|Z|YD IP (X, Y, Z).
Suppose further that if D were to lose any of its arrows, then this implication
would no longer hold (i.e. suppose that D is minimal in this regard). Under these
suppositions, D is said to be a Bayesian network of the probability distribution P
(on the nodes of D).6
The question now arises: given a probability distribution P, how might one
determine a Bayesian network of P? That is, how might one find a directed
acyclic graph (for short: a dag) D, using the nodes on which P is defined, so
as to ensure that
X|Z|YD IP (X, Y, Z),
and also that D is minimal in this regard? The answer is that, relative to any
(linear) ordering < of the nodes, one can define the notion of a boundary dag
for P relative to <. (The details will not detain us here.7 )
304
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
It can be shown8 that for any probability distribution P, the ternary indifference relation IP (X, Y, Z) satisfies four simple conditions.9 One can then prove
that it follows from the fact that IP (X, Y, Z) satisfies these four conditions that
any boundary dag for P relative to any ordering of the nodes is a Bayesian
network of P. 10
Pearls main theoretical aim is to supply a computationally efficient method
for propagating changes in probability values along the links in a Bayesian
network, so that a new equilibrium will be reached in response to the initial
changes. To this end, he pays attention to the kinds of structure that a network
can exhibit. In broad terms, he is seeking to canalize or confine the effects
of such changes, 11 so that they propagate in as linear a fashion as possible,
and by means of algorithmic calculations that are as local as possible. (The
reader will recall in this connection our earlier discussion of greedy algorithms
in Section ..) The aim is to cope with feedback loops, and ensure that at
equilibrium, the correct probabilities will be assigned to every node. 12
This brief exposition will have to suffice, for we are concerned here only to
make a few important points of contrast. It is clear that, as Pearl conceives his
project, the nodes can represent events that can be part of a causal network. But,
as he has clarified further (personal communication),
(Clearly, B = . So the node X will have no parents in the sought boundary dag.) At each ith
stage, we are creating Bi by selecting as few nodes as possible from among X , . . . , Xi , thereby
leaving as many nodes as possible in the residue
Bi = {X , . . . , Xi } \ Bi ,
in such a way as to satisfy the independence condition
IP (Xi , Bi , Bi ).
We repeat that the sought boundary dag is obtained by making the nodes in Bi the parents of the
node Xi .
See Pearls Theorem , p. .
The conditions in question are as follows (ibid., at p. ):
1.
2.
3.
4.
B AY E S I A N N E T WO R K S
305
There is no restriction that the links represent causal relations (although this has many
advantages). Belief propagation will work as well on non-causal networks, be it probabilistic
or logical, as long as the graph is compatible with the underlying probability function.
So the arrows in Pearls belief nets could represent the lending of justificatory
support, whether singly or in larger sets. And presumably this could be so
even in the extreme case where the only probability values in use are and
(corresponding to a completely deterministic picture, in the causal case). So his
belief nets could conceivably deal with, say, mathematical beliefs as a special
case. Despite the fact that their original exposition seemed to be concerned
with only the causal caseinvolving simple contingent propositions about the
external world, all of them caught up in a causal webwe have to take seriously
that Pearls belief nets have as wide an intended range of application as do our
dependency networks.
But then, from our own formal modeling perspective, special interest
attaches to the aforementioned lack of any use of inference strokes to mediate
the connections among nodes. Because Pearl does not employ any such device,
it might appear there is no way of telling whether the arrow-fork
a
represents a and b conjointly (i.e. together) causing c, or represents the conjunction of as causing c along with bs (independently) causing c. As Pearl has
made clear, however (personal communication), In belief nets, the distinction
you are making in the two graphs [see belowNT] is encoded in the probability
tables, not in the graphs.
This raises two questions worthy of further investigation:
(i) Is there an algorithmic way to read off , from the probability tables, the appropriate placement of inference strokes and arrows for the theorist of dependency
networks?
(ii) Might the placement of inference strokes and arrows by the theorist of dependency networks be a fruitful additional device for the theorist of belief nets?
In any future extension of the present work to the probabilistic case where intermediate probability values are countenanced (rather than just the two extreme
306
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
values of and ), we would be concerned to investigate what theoretical benefits might flow from using this apparently innocent device (i.e. the inference
stroke) to make the necessary (and very useful) distinction:
a
versus
c
Prob(a)=p1
Prob(b|a)=q1
Prob(b|a)=q2
B AY E S I A N N E T WO R K S
307
the dependency network theorist would separate out the values a and a, and
supply them with their own nodes. The conditional probability values could
then be transferred to the inference strokes in order to register the respective
strengths of support:
b
q1
a
p1
q2
a
1p1
308
D I F F E R E N C E S F R O M OT H E R F O R M A L T H E O R I E S
C H A P T E R 12
Our interest in the works thus classified is accordingly motivated by the following questions, respectively.
S TAG E S E T T I N G
309
1. (a) Do these works lay out the general features of belief revision in a way that
is at least broadly in accord with the theory advanced here?
(b) Do they set about the task of formal modeling in ways that are at odds
with our own chosen methods, or in ways that could be emulated by (some
plausible extension of) them?
2. (a) as above.
(b) Could the formal modeling that we offer be plugged into those works, so
as to fill understandable, but theoretically hospitable, lacunae?
3. Do these works paint an informal picture of the structure of our beliefs, and
possibly also of their revision, that offers at least some prospect of being
captured by our framework?
epistemology
12.2.1 Skepticism
Evil demons or brains in vats?we dont go there. Nor, apparently, are we
alone in this regard. According to Hendricks [], formal epistemologies are
310
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
marked, one and all, by the distinguishing feature that they build in some sort
of assumption to the effect that radical skepticism does not hold.
On the other hand, there is nothing to stop us from using our formal modeling of belief systems to represent the beliefs of a philosopher grappling with the
problem of radical skepticism and eventually deciding to resolve the problem one
way or another. Happily, the construction of suitably accommodating before
and after dependency networks can be left to the reader who insists that we
ought to be able to model the processes of belief change on the part of an
agent in the grips of this, the deepest of all skeptical problems. In exactly the
same way, one supposes, formal logic can be put at the disposal of someone
wishing to construct her horoscope or read the future in her tea leaves. Good
luck constructing all the steps that have (either as a premise or as the conclusion)
the node for I am a brain in a vat! It might bring some clarity to the debate about
the issues.
311
And the same holds once more with regard to the various answers on offer
to the question To what does a basic belief owe its justification?. Its being basic
consists at least in this: a basic justified belief does not owe its justification to
any other beliefs; and a basic justified belief is one whose justification does not
depend on any justification the agent has for holding any other belief. Both these
features of basicality are nicely accommodated by the role allotted to initial
nodes within our dependency networks.
The much-cited paper Sosa [] is famous for its contrasting metaphors for
coherentism and foundationalism. Here we concentrate on what Sosa says about
foundationalism. The definition that Sosa provides on p. is as follows.
With respect to a body of knowledge K (in someones possession), foundationalism implies
that K can be divided into parts K1 , K2 ,, such that there is some nonsymmetric relation
R (analogous to the relation of physical support) which orders those parts in such a way
that there is onecall it Fthat bears R to every other part while none of them bears
R in turn to F.
According to foundationalism, each piece of knowledge lies on a pyramid such as the
following:
P
P1
P11
P2
P12
P21
P22
The nodes of such a pyramid (for a proposition P relative to a subject S and a time t)
must obey the following requirements:
a. The set of all nodes that succeed (directly) any given node must serve jointly as
a base that properly supports that node (for S at t).
b. Each node must be a proposition that S is justified in believing at t.
c. If a node is not self-evident (for S at t), it must have successors (that serve jointly
as a base that properly supports that node).
d. Each branch of an epistemic pyramid must terminate.
In order to secure the right picture, Sosa needed to say that the relation R is nonsymmetric and transitivefor otherwise it would be easy to construct systems
312
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
For the mathematician or computer scientist these definitions are at best suggestive. The more formally minded theorist would want, ideally, a definition of
coherentism that would make it clear just what sorts of patterns of justification
(involving premises and conclusions of justificatory steps) are allowed, and
which are disallowed, by the coherentist.
P E R E N N I A L TO P I C S O F M A I N S T R E A M E P I S T E M O L O G Y
313
This appears promising: an epistemologist apparently undertaking to give a precise account of those global patterns within the field of the justificatory relation
that are characteristic of coherentism. It is a heavy obligation to discharge; for
(p. )
the justification of a particular empirical belief finally depends, not on other particular
beliefs as the linear concept of justification would have it, but instead on the overall system
and its coherence.
So: the reader is left with the impression that a defence is to be mounted of a view
that, it would appear, need not be characterized at all precisely. But BonJour does
provide some hints (pp. ):
Intuitively, coherence is a matter of how well a body of beliefs hangs together: how well
its component beliefs fit together, agree or dovetail with each other, so as to produce an
organized, tightly structured system of beliefs, rather than either a helter-skelter collection
or a set of conflicting subsystems. It is reasonably clear that this hanging together depends
on the various sorts of inferential, evidential, and explanatory relations which obtain among
the various beliefs, and especially on the more holistic and systematic of these. Thus
various detailed investigations by philosophers and logicians of such topics as explanation,
confirmation, probability, and so on, may be reasonably taken to provide some of the
ingredients for a general account of coherence. But the main work of giving such an
account, and in particular one which will provide some relatively clear basis for comparative
assessments of coherence, has scarcely begun, despite the long history of the concept.
My response to this problem, for the moment at least, is a deliberatethough, I think,
justifiedevasion. It consists in pointing out that the task of giving an adequate explication
of the concept of coherence is not uniquely or even primarily the job of coherence
theories. This is so because coherenceor something resembling it so closely as to be
314
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
subject to the same sort of problemis, and seemingly must be, a basic ingredient of
virtually all rival epistemological theories as well.
The rest of BonJours discussion in this section (Section .), titled The concept
of coherence, dwells on matters such as logical and probabilistic consistency,
and explanatory unification, without providing any of the sorts of structural
insights that the formal theorist would be seeking.
Nor has the passage of time, for the coherentists, brought greater clarity
at a formal level. The following quote from an authoritative survey article by
Kvanvig [] is also typical, and illustrates the same problem (of structural
unclarity) just as well:
Sometimes coherentism is described as the view that allows that justification can proceed
in a circle (as long as the circle is large enough), and that is one logically possible version
of the view (though it is very hard to find a defender of this version of coherentism). The
version of coherentism that is more popular, however, objects in a more fundamental way
to the regress argument. This version of coherentism denies that justification is linear in the
way presupposed by the regress argument. Instead, such versions of coherentism maintain
that justification is holistic in character, and the standard metaphors for coherentism are
intended to convey this aspect of the view. Neuraths boat metaphoraccording to which
our ship of beliefs is at sea, requiring the ongoing replacement of whatever parts are
defective in order to remain seaworthyand Quines web of belief metaphoraccording
to which our beliefs form an interconnected web in which the structure hangs or falls as
a wholeboth convey the idea that justification is a feature of a system of beliefs.
From the point of view of the formal modeler, learning that according to coherentism justification is a feature of a system of beliefs is not terribly helpful.
It hardly speaks to the question of what sort of structure of justification the
coherentist allows.
Kvanvig envisages a form of coherentism in which one can have recourse to
experiences (without propositional content) as playing some role in the overall
justification of an agents beliefs. It is important for the coherentist only to
(i) reckon any and all experience as insufficient for full justification of the whole
(so as to avoid collapsing into a form of foundationalism); and (ii) recognize
the important complementary role played, for any given belief, by its evidential
and logical relations to other beliefs in the system.
In our formal modeling, modest as it is, we seek to accommodate virtually
any constraining (or generating) of structural possibilities by a suitably rigorous
characterization of coherentism, should one ever be forthcoming from either its
advocates or its critics. We expressly allow for there to be nodes within a dependency network that are both initial or self-warrantingthat is, representing
P E R E N N I A L TO P I C S O F M A I N S T R E A M E P I S T E M O L O G Y
315
beliefs that are believed outrightand justified by appeal to yet other beliefs
functioning as premises in transitional steps with those beliefs as conclusions.
This allows one to accommodate both (i) and (ii) above, in whatever proportions they might find themselves in the right coherentist admixture. Although
Grdenfors [] has argued that AGM-theory is a belief-revision theory suitable for a coherentist epistemology, this claim has been disputed by Hansson
and Olsson []. Fortunately we do not need to take a stand on this issue.
We could instead invite the two sides to that debate to avail themselves of the
formalism of dependency networks in order to capture the essential features
either of foundationalism or of coherentism, howsoever they (especially the
latter) might end up being articulated.
FOUNDHERENTISM
Susan Haacks Evidence and Inquiry (Haack []) advances an interesting and
detailed case for foundherentism, a hybrid view that combines the virtues of
both foundationalism and coherentism. The main ingredients might have been
anticipated by Cornman [].2 But credit goes to Haack for the first booklength exposition of the view. Quite apart from certain drawbacks in Cornmans
account (see below), Haacks more incisive analyses and livelier style invite
scrutiny of the foundherentists position as she has laid it out.
For Haack, as for many other epistemologists (including Cornman and Sosa),
justification of a belief by evidence is relative to a person and a time, and comes
in degrees. (This mention of Sosa in this context is not to be understood as
implying that Sosa himself is a foundherentist.) The formally minded theorist,
however, will be disappointed by Cornmans definition, on p. , of the crucial
notion of an evidential series for a proposition p. For Cornman, this is a
series [E] of evidence sets, e1 , e2 , e3 , . . . , such that for each ei in E and e0 = p, each
sentence in ei1 is justified relative to ei . . . .
With this definition, Cornman erases, or loses sight of, the crucial branching structure of justification, which would be the basic ingredient in terms
of which to account, in a more subtle and insightful way, for the differences
between foundationalism and coherentism at the global level. If, for example, the justification of p had the following structure (in our diagramming
notation):
316
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
317
t
..
.
v
..
.
w
..
.
x
..
.
y
..
.
..
.
..
.
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
319
Although Levi emphasizes a distinction between coerced and uncoerced contractions (p. ), it does not affect the mathematical constructs or methods
involved in the contraction process. So Levis account inherits the drawbacks of
the AGM account. The theories (or belief sets) that he deals with are infinitistic.
There is no guarantee that one can manipulate them computationally. This, of
course, is not meant to imply that Levi himself had entertained any aspirations
at all to be able to handle his models computationally. His interest lay rather
in how best one might formally model the full belief commitments of a rational
agent.
Now, to be sure, this is not by itself a telling criticism of Levis project, for
Levi, unlike the AGMers, has not claimed on behalf of his theory that it would
be of interest and use in the fields of computing science or AI. In raising the
issue of constructivity or computability we are merely seeking to emphasize
an important difference between Levis approach and ours. One would have to
judge on more global, pragmatist criteria whether a theory such as ours, which
seeks to keep everything computational, is to be preferred to a theory such as
Levis, which does not. In assessing this matter, it would be important to check
whether there is any intuitive example raised by Levi that cannot be handled
straightforwardly in the framework of finite dependency networks. If there are
no such examples, then there is a prima facie reason to prefer the more modest
theory over one that postulates infinitistic objects and invokes structures such
as probabilistic distributions over them.
The main innovation that Levi introducesa measure of so-called damped
informational valueis not of a constructive or computable character either.
Thus, when Levi finally offers his considered choice of contraction (p. ):
K
A (the admissible contraction of K by removing A) is the meet of all saturatable
contractions of K removing A that minimize loss of informational value among saturatable
contractions as determined by the given M function
the computationalist finds no succour in such slight variation of basically AGMstyle thinking, in which belief sets are treated as logically closed.
The M function is defined (p. ) as a probability distribution over belief
sets K (which, remember, are logically closed); and from M one defines a
measure of informational value in some inverse fashion. Three possibilities that
Levi mentions are M(K), /M(K) and log[M(K)], and he expresses a
preference for the first. But there is no hint there, either, of how to compute
these informational values from their arguments K (which cannot, in general,
be presented finitistically anyway).
320
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
Levi is not, however, detained by the question whether one is entitled to assume
the existence of such an underlying probability measure M as would vouchsafe
the satisfaction of this assumption. For all we know, the assumption just cited
could be inconsistent with the Kolmogorov axioms for probability. At the very
least, one would need some assurance to the contrary.
The prima facie problem is that informational value is said by Levi to be an
inverse function of probability, and the three mooted examples of such inverse
functions are all monotonic. Hence equality of losses of informational value
implies equality of the respective probabilities from which the informational
values are supposedly derived. This means (in light of the last quote above)
that the probability of an intersection of finitely many (saturatable) contractions
must be equal to the lowest probability of those finitely many contractions.
There is no guarantee that any of those finitely many contractions will be identical to their intersection. Hence any proposition in any contraction of lowest
probability among the contractions being intersected that does not lie in the
intersection in question must be assigned the informational value , hence
the probability value . The worry is not only that this cannot in general be
guaranteed, but that it would appear to be rather irrational to be giving up
(potentially very many) beliefs with prior probabilities of .
In Levi [], pp. , Levi modifies the foregoing assumption, but his
modification attracts the same misgiving by way of criticism. The assumption
now reads
any contraction that is the intersection of finitely many maxichoice contractions (of a belief
set with respect to a given sentence) incurs a loss of (damped) informational value equal
to the greatest such loss incurred by any of the finitely many contractions in question.
321
the one about tainted hummus, discussed in Section .., receive intuitively
correct treatment with our method of dependency networks, and our contraction algorithm. So the continued resort to infinitistic theories, replete with conjectured probability- and associated informational-value measures, strikes the
present author as an unnecessary formal complication for the computationalist.
This runs against the foundationalist grain, and is a distinctive mark of a coherentist view.
What distinguishes their discussion is their treatment of the five virtues
of scientific hypotheses: conservatism; modesty; simplicity; generality; and
refutability. Their analysis both of examples and of the general nature of these
virtues reveals both the role of logic in belief statics and some methodological
322
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
principles of belief dynamics. Two tenets from the latter are reiterated in various
places:
(i) when a set of beliefs is found to be inconsistent, there is a plurality of choices
of particular beliefs to surrender or reject; and
(ii) every belief can be made vulnerable to surrender or rejection, provided only
that one is prepared to make suitable adjustments elsewhere within ones belief
scheme.
Tenet (i) is what underlies the non-deterministic character of our algorithm for
contraction.
In our language, tenet (ii) would read: it is always possible to find a contraction of ones belief scheme, no matter which node within it is initially whitened.
We have seen also in Section .. how Quine and Ullians detailed example
of revision of beliefs about the identity of the murderer can be handled by our
account, using finitary dependency networks.
All in all, The Web of Belief sets out an epistemological account that is completely amenable to modeling by our formal methods, without violating any of
Quine and Ullians fundamental precepts.
323
dency network only those beliefs that are focused and informative, and not
diluted by, say, unnecessary steps of disjunction introduction. But what Harman
gives with the right hand he takes away with the left. He claims that one can
and does believe infinitely many things, and gives as examples (for values of
n greater than ) sentences of the form the earth does not have n suns. So, in
order to accommodate this point, he restricts the Principle of Clutter Avoidance
to apply to what one believes explicitly.
This brings in its wake three distinctions among beliefs (which it is our task to
represent as nodes in the agents dependency network): explicit v. implicit; conscious v. unconscious; and occurrent v. dispositional. From our point of view,
the only interest in drawing these distinctions is to be aware of the difficulties
one might face in teasing out (on behalf of a rational agent) exactly what the
dependency network for her beliefs really is. Once we have the dependency
network in hand, these distinctions have no further normative bearing on the
ways in which she ought to change her mind when the occasion arises. Appeal
to these distinctions might be useful in order to explain shortfalls in an actual
(unidealized) agents performance when she tries, unsuccessfully, to make effective changes; but the distinctions do not bear on matters of competence (on the
part of an idealized agent) as opposed to performance, once the dependency
network has been identified.
Perhaps the single most important claim of Harmans that is inimical to the
project presented here is to be found on p. , where he writes that
although in revising ones view it would be useful to have a record of ones reasons for
various beliefs, this would involve too much record keeping.
In the alleged absence of adequate records of ones reasons for various beliefs,
he suggests that
one must make do with a principle of positive undermining, which takes any current belief
to be justified as long as there is no special reason to give it up.
If one really did keep no mental records at all of the reasons for ones beliefs,
then the dependency-network account on offer here would be over-ambitious
at best, and farcical at worst. There would just be a collection of nodes, with
no connections among them: no inference strokes, and no arrows (as in our
diagrams of them). This, however, is an unrealistic picture of the familiar phenomenon. Faced with strong evidence against ones former belief that p, most
rational agents can think of all sorts of justificatory connections, both to and
from p, that had moved them in their earlier thinking that led them to p.
324
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
325
The lines of dependency of derived beliefs on the more immediate are certainly needed by a paragon, or rational agent. The question concerns only the
extent to which it would be the paragons richer dependency network, or the
rattling belief box, that provided the more accurate model of what goes on in
the mind of a reasonably thoughtful person. The contention here is that it is the
dependency network. It is an empirical question, to be sure.
We contend further that the likeliest empirical answerat least concerning
really reflective and thoughtful people, who are well trained in the weighing
of evidence and the drawing of inferencesis likely to be encouraging to the
dependency-network theorist. The answerconcerning the justificatory pedigrees of the beliefs held by the best experts among usis likely to be one that is
easily generalized and strengthened so as to make the dependency network of
the paragon the most reasonable ideal to which we should aspire if we wish to
be truly rational, and accountable by our own standards, when we contract and
expand our belief systems. Similar qualifications have to be entered on behalf of,
say, the rules of natural deduction and the structure of proofs in mathematics.
Not only are most ordinary thinkers unacquainted with formal deductive rules
of inference; they commit many fallacies of deductive reasoning, and they even
fail at times to appreciate that certain formally correct steps of inference are
indeed correct. But the best-trained experts among uslogicians and mathematicians, for examplereason deductively (and respond to proofs) in a way
that makes the best description of their performance invoke rules drawn from
the normative canon.
A related claim of Harmans with which we take issue is the following:
We could not in practice make use of a principle of negative undermining, which has one
abandon any belief for which one does not have a current justification. Following that
principle would lead one to abandon almost everything one believes.
Again, it is the bleak picture of a rattling belief box, filled with unconnected
nodes, that is at work here. If Harman is wrong (empirically) about the extent to
which thoughtful people descry justificatory connections among their beliefs,
then they should be beholden to a principle of negative undermining. This is
especially the case if we say that one should abandon any belief for which we
would not have any present justification in terms of justificatory relations that one
has brought to bear in the past. On Harmans pessimistic and opposing picture,
the downward sweeps of our contraction algorithm would not whiten any nodes
that had not already been whitened in an earlier upward sweep, as members of
an occluder.
326
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
Harman also has a rather unusual view (p. ) connected with his refusal to
set any store by justificatory links among beliefs (or the agents potential access
to them as needed). He characterizes the distinction between foundationalism
and coherentism in the following rather unorthodox way:
The key issue is whether one needs to keep track of ones original justifications for beliefs.
What I am calling the foundations theory says yes; what I am calling the coherence theory
says no.
By contrast with Harman, on our account of the matter, it is not that the foundationalist does, while the coherentist does not, acknowledge and make use
of justificatory relations among beliefs when changing ones mind. Rather, the
distinction between the two is to be found in the overall patterns of justification
that are permitted or prohibitedpatterns that very definitely need justificatory
relations to be in place, so that their overall structural features can be read off.
On our account, the intuitive idea is the right one: the foundationalist sees justifications as tree-like in structure (or like chicken wire,4 with all arrows tending
downwards), whereas the coherentist sees justifications as web-liketangled
skeins, allowing loops. The great advantage of our account is that our method
of contraction (and therefore also of revision) is indifferent to the distinction
between the two positions, and is invariant across them. One and the same
method works for both.
On some other important matters, our account accords with Harmans. He
maintains that one has only a finite number of explicit beliefs (p. ). He also
takes belief to be an all-or-nothing matter (p. ). This, however, is because he
thinks that if beliefs are represented as having degrees (or probablility values),
then the task of updating by conditionalization would be too complicated in
practice. By contrast, we take belief to be an all-or-nothing affair only because
we are concerned to keep matters as simple as possible in this, our first detailed
attempt to model a difficult subject matter. Moreover, Harman is making an
intuitive judgement about complication (or complexity) not backed by such
formal work as would be necessary in order to justify it. And that summary
judgement would appear to be undermined by such formal work on probabilistic belief nets as has been done in this area, such as Pearl []. Harman claims
Todays politically correct synonym is poultry netting, which one presumes is intended to avoid
giving any offence to turkeys, ducks, geese or swansor to any breeders thereof, who might have
suffered twinges of guilt over the impropriety of bending bits of chicken wire in their husbandry of
these other species.
327
Note the crucial relativization for A. This is in agreement with our account.
The main difference is that we seek to work with these basic materials, in order
to give a richer framework of norms for belief change. Moreover, we investigate
just how computationally complex the process of changing ones mind (i.e. ones
belief system) actually is. And those results go against the general tenor of
Harmans account, which is a curious combination of empirical pessimism and
theoretical hesitancy. In making his account hostage to empirical fortune, he
will be left (if the empirical facts turn out to be disconfirming) with no normative account of how a rational agent ought to change her beliefs.
328
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
The intended reading, of course, is that for i < j, we have ri = rj . For Klein is
explicitly envisaging the possibility of infinitely regressing chains of reasons for
Ss belief that x. The chain of reasons, he writes (for any belief), cannot end with
an arbitrary reasonone for which there is no further reason. And in meeting
this requirement, one cannot have recourse to coherentist loops, since these
have been ruled out (p. ) by the Principle of Avoiding Circularity (PAC):
For all x, if a person, S, has a justification for x, then for all y, if y is in the evidential ancestry
of x for S, then x is not in the evidential ancestry of y for S.
The combined effect of PAC and PAA is that for every belief node x for a rational
agent S, there will be an infinite partial ordering of yet further nodes with x as
its minimum. The node x will be less than each node in any premise set of other
nodes offered in support of x. Any node x in any such premise set will in turn
be less than each node in any premise set of other nodes offered in support of
x ; and so on. This infinite partial ordering with x as its minimum need not be
a tree, since a justificatory ancestor y of x, lying above x in this ordering, might
be reached by following two distinct paths generated by the less than relation:
Klein makes it clear that he is not envisaging an infinity of actually entertained
reasons regressing upwards in the way indicated in our diagram. His contention
(interpreted in terms of our diagrams) is only that no node in the diagram should
ever be taken as initial.
THE PROBLEM OF BELIEF REVISION
329
..
.
..
.
..
.
..
.
..
.
y
..
.
..
.
..
.
..
.
..
.
Let x be any node. Then, according to Klein, x would have an inference stroke
above it, and the agent might5 be able to find at least one more belief nodeone
that is both subjectively and objectively availablefrom which to supply an
arrow extending down to that inference stroke. Call such a further belief node w.
Klein does not claim that the agent need ever have considered or entertained
w at any time before acquiring the justified belief x. He carefully distinguishes
(following Fumertonsee p. ) between xs having a justification proceeding
via w, and xs being justified (for the agent) in virtue of a train of past justificatory
thought proceeding via w.
This concession by Klein, it seems to the present author, takes the infinitist
sting out of any objection to our finitary approach to contraction and revision.
It is all that we need seize upon in order to avert any principled objection by the
infinitist to our accounting for contraction (and revision) in terms of finitary
dependency networks. Ours has always been an agent-relative, though normative, account. It holds the agent accountable to the steps that she has recognized,
and the justifications that she has put together for her various finitely many,
actually formed, beliefs. So, confronted with a potentially infinite Kleinian tree
of justifications that a resourceful agent might be disposed to generate upon
challengeeven if such justifications had never been entertained before forming the beliefs in questionwe can simply prune it down by considering only
those nodes that have played a justification-transmitting role for the agent. The
As Klein has clarified (personal communication), certainly there will be times when at least
for a while (a good long whilelonger than the agents lifetime) the agent wont be able to locate
the next highest node.
330
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
result of such pruning, Klein concedes, is finite. In the language of Klein [],
the result consists of the doxastic justifications possessed by the agent, regardless
of how infinitely various and infinitely regressive the so-called propositional
justifications might be. The infinitist will take the belief that p to be doxastically
justified for S just in case S has engaged in providing enough reasons along an
endless path of reasons. (Loc. cit., p. .) Presumably what counts as enough
here will always be finite. In more detail (p. ):
Infinitism is committed to an account of propositional justification such that a proposition,
p, is justified for S iff there is an endless series of non-repeating propositions available to
S such that beginning with p, each succeeding member is a reason for the immediately
preceding one. It is committed to an account of doxastic justification such that a belief
is doxastically justified for S iff S has engaged in tracing the reasons in virtue of which
the proposition p is justified far forward enough to satisfy the contextually determined
requirements.
Actual tracing takes time; so the set of beliefs that enjoy doxastic justification
for the agent is finite.
It follows that we shall always be able to keep abreast with the agent who
is re-thinking any matter, since, ex hypothesi, the agent will only ever have
generated but finitely many of those nodes that Klein imagines as crystallizing
out indefinitely at ever-higher levels within a potentially infinite diagram.
331
framework, it would fit in; but he does not provide any details. It is as though
he is telling his reader where exactly to put the black box, but is not venturing
to open it, or even speculate about its internal structure.
The following passage (on p. ) in effect extends an invitation to supply a
workable account of the belief-revision process. We cautiously offer the present
account (or some appropriate extension of it) as a plug-in for Guptas epistemology:
Suppose that an experience e yields, when conjoined with a view v, a class of judgments
that contains Q. Let us represent this as follows:
(2) Q e (v).
Suppose further that we have a rational being who holds the view v and suffers the
experience e. This being has to cope with the rational constraint expressed in (2). In most
cases, the constraint is met simply by accepting Q, but in some cases this may require
a modification of the view v. I shall sometimes read (2) as saying that the experience e
and the view v entail the judgment Q.[fn] The present point, then, is that the character of
this entailment is similar to that of logical entailment: it does not always yield entitlement.
A rational being that holds the view v and suffers the experience e is not automatically
entitled by (2) to the perceptual judgment Q. Constraint (2) forces the rational being
to adjust its view v in light of Q (and the other judgments in e (v)). Most often this is
achieved simply by adding Q to v. But sometimes it requires a substantial revision of the
original view v to a new view v . It is possible that v does not sustain the judgment Qit
may even be that e and v entail the negation of Q. In such a case, the total effect of e
and v precludes entitlement to Q.
C O N N E C T I O N S W I T H E P I S T E M O L O G I C A L AC C O U N T S
the usual accounts of revision. Yet there is an asymmetry of sorts already built
in, on Guptas approach, since the set J is of the form e (K) (taking K as the
prior view v). And by making J depend thus on K, one would have thought
that the -function would somehow have recourse to K in order to ensure that
its output e (K) would not be so contrary to K that it could turn out that no
part of J = e (K) need survive in K J.
Does Gupta really want this extreme possibility to remain a possibility? Or
can the recency of e be held to be a consideration in favor of retaining at least
some of e (K) within the revision K e (K)?
If the answer to this last question is negative, as Gupta thinks it is, then
what we can offer Gupta is only our account of inconsistency-eliminating
contractionwhich is all that he would be seeking, anyway. On this account
our aim is to whiten absurdity () at the initial stage, and then to perform
upward and downward sweeps according to the algorithm(s) we have provided.
In performing such consistency-restoring contractions, no great store would
be set by e (K), despite its recency (for the rational agent in question).
If, on the other hand, the answer to our last question is affirmative, then our
account of contraction can be applied in full. We would need only to have some
non-empty but proper subset I of e (K) chosen for initial whitening; and the
contraction algorithm could then be applied in the usual way so as to compute
contractions of K (| e (K)) with respect to I. (Here the notation | U, for U a
set of nodes, is the set of all initial beliefs of the form | u, for u U.)
Either way, it would seem that our account (or, if necessary, some extension
of it that would allow for defeaters) would be the appropriate filling-in of the
details of the black box whose position within Guptas overall account is made
so clear by the foregoing quote from Empiricism and Experience.
333
REFERENCES
Carlos Alchourrn and David Makinson. On the logic of theory change: contraction
functions and their associated revision functions. Theoria, :, .
Carlos Alchourrn and David Makinson. On the logic of theory change: Safe contraction. Studia Logica, :, .
Carlos Alchourrn, Peter Grdenfors, and David Makinson. On the logic of theory
change: partial meet contractions and revision functions. Journal of Symbolic Logic,
:, .
Alan Ross Anderson and Nuel D. Belnap. Entailment: The Logic of Relevance and Necessity. Vol. I. Princeton University Press, Princeton, NJ, .
Robert Audi. The Architecture of Reason: The Structure and Substance of Rationality.
Oxford University Press, Oxford, .
Alfred J. Ayer. Language, Truth and Logic, nd edn. Gollancz, London, .
David Bayles and Ted Orland. Art & Fear: Observations On The Perils (and Rewards) of
Artmaking. The Image Continuum, Santa Cruz, CA & Eugene, OR, .
Laurence BonJour. The Structure of Empirical Knowledge. Harvard University Press,
Cambridge MA, .
Alonzo Church. A note on the Entscheidungsproblem. Journal of Symbolic Logic,
:, a.
Alonzo Church. Correction. Journal of Symbolic Logic, :, b.
Alonzo Church. Review of Ayers language, truth and logic, nd edition. Journal of Symbolic Logic, ():, .
S. A. Cook. The complexity of theorem-proving procedures. Proc. rd Annual ACM
Symposium on Theory of Computing, pages , .
James W. Cornman. Skepticism, Justification, and Explanation: with a Bibliographic Essay
by Walter N. Gregory. Reidel, Dordrecht, .
Donald Davidson. Actions, Reasons, and Causes. The Journal of Philosophy, LX():
, .
REFERENCES
335
336
REFERENCES
REFERENCES
337
338
REFERENCES
Neil Tennant. Anti-Realism and Logic: Truth as Eternal. Clarendon Library of Logic and
Philosophy, Oxford University Press, Oxford b.
Neil Tennant. Autologic. Edinburgh University Press, Edinburgh, .
Neil Tennant. Changing the Theory of Theory Change: Towards a Computational
Approach. British Journal for Philosophy of Science, ():, a.
Neil Tennant. Intuitionistic Mathematics Does Not Need Ex Falso Quodlibet. Topoi,
pages , b.
Neil Tennant. Changing the Theory of Theory-Change: Reply to my Critics. British
Journal for Philosophy of Science, :, a.
Neil Tennant. On Having Bad Contractions: or, No Room for Recovery. Journal of
Applied Non-Classical Logics, :, b.
Neil Tennant. The Taming of The True. Oxford University Press, c.
Neil Tennant. Negation, Absurdity and Contrariety. In Dov Gabbay and Heinrich
Wansing, editors, What is Negation?, pages . Kluwer, Dordrecht, .
Neil Tennant. Deductive v. Expressive Power: A Pre-Gdelian Predicament. Journal of
Philosophy, XCVII():, .
Neil Tennant. Theory-Contraction is NP-Complete. The Logic Journal of the IGPL, ():
, .
Neil Tennant. Contracting Intuitionistic Theories. Studia Logica, :, .
Neil Tennant. New Foundations for a Relational Theory of Theory-Revision. Journal of
Philosophical Logic, ():, a.
Neil Tennant. A Degeneracy Theorem for the Full AGM-Theory of Theory-Revision.
Journal of Symbolic Logic, ():, b.
Neil Tennant. Belief-Revision, the Ramsey Test, Monotonicity, and the so-called Impossibility Results. Review of Symbolic Logic, ():, .
Neil Tennant. Cut for Core Logic. Review of Symbolic Logic, forthcoming.
Paul Thagard. Computational Philosophy of Science. MIT Press, Cambridge, MA,
.
Paul Thistlewaite, Michael A. MacRobbie, and Robert K. Meyer. Automated TheoremProving in Non-Classical Logics. Research Notes in Theoretical Computer Science.
Pitman, London and Wiley, New York, .
J. Ullian. A Note on Scheffler on Nidditch. Journal of Philosophy, :, .
Alasdair Urquhart. The undecidability of entailment and relevant implication. Journal of
Symbolic Logic, :, .
Alasdair Urquhart. The Complexity of Decision Procedures in Relevance Logic. In
J. Michael Dunn and Anil Gupta, editors, Truth or Consequences: Essays in Honor
of Nuel Belnap, pages . Kluwer, Dordrecht, .
Iris van Rooij. The Tractable Cognition Thesis. Cognitive Science: A Multidisciplinary
Journal, ():, .
REFERENCES
339
T. S. Verma. Causal networks: Semantics and expressiveness. Technical Report R-. Cognitive Systems Laboratory, University of California, Los Angeles, .
Florian von Schilcher and Neil Tennant. Philosophy, Evolution and Human Nature. Routledge & Kegan Paul, London, .
340
REFERENCES
INDEX
<-violation,
LR,
AGM-theory, viii, ix, , ,
, , , , , ,
, ,
laxity of,
Absurdity Rule,
CR,
DNE,
EFQ,
Ex Falso Quodlibet,
IR,
perfection constraint,
,
(honesty), ,
(success), ,
closure,
consistency constraint,
,
inclusion,
macro-structural
constraint,
minimal mutilation, , ,
,
preservation,
recovery,
success, , , ,
vacuity,
absolutely bizarre revisions,
absolutist,
absurdity constant,
Action Types, , ,
Action Type ,
Action Type ,
Action Type ,
Action Type ,
admissible contraction,
adopting, , , , , ,
, ,
agent-relativity, ,
Alchourrn, Carlos, viii, ,
,
algorithm
exponential-time,
for existential closure,
greedy, , , ,
,
non-deterministic, , ,
, ,
polynomial-time,
staining, ,
Anderson, Alan Ross,
,
Arl-Costa, Horacio, xi
Arora, Anish, xi,
assumption-retraction,
Axioms of Coloration, , ,
, , , , ,
, ,
Axioms of Configuration,
, , , , ,
Ayer, Alfred J., viii, ,
baffle, , ,
bafflement,
base
irredundant, , ,
basic beliefs, ,
Bayesian networks, , ,
, , ,
belief nets, ,
belief scheme, ,
belief set, , , , ,
, , , , ,
Belnap, Nuel, , ,
binary search,
bloating,
maximal,
minimal, ,
BonJour, Laurence, , ,
Buss, Sam, xi
Carnap, Rudolf,
Carter, Eric, xi
casualty, , ,
categorical description, ,
Chain, , ,
chains, infinitely descending,
Chandrasekaran, B., xi
Churchs Theorem,
Church, Alonzo, viii,
Circle, , ,
INDEX
341
classical propositional
calculus,
Classical Reductio ad
Absurdum,
closure
T-closure, , , ,
,
downwards,
existential, , , ,
, , ,
existential, algorithm for,
logical, , , , , , ,
, , , ,
under known steps, , ,
, ,
universal, , , ,
, ,
upwards,
cluster, , , ,
, , , ,
Clutter Avoidance, ,
co-NPNP -complete,
co-NP-complete,
cognitive significance,
coherentism, , , , ,
,
Cole, Julian, xi
collapse
of indirect verifiability, ix
of AGM-theory, viii
collapse proof,
coloration of diagram, , ,
, , , , , , ,
, , , , ,
, , ,
competence, vii
completeness,
components of stage of
computation,
comprehensiveness,
computational complexity, ,
, ,
computational saint, ,
condition (M< ),
conjunctive normal form,
connectives,
conservatism,
constraint (),
contingent sentence,
proposition or belief,
342
INDEX
, , , , ,
,
contraction, , , , ,
, ,
algorithm, , , , ,
, , , , , , ,
, , , , ,
, , , ,
, , ,
, , ,
, , , , ,
, , , , ,
, , ,
clear intuitions about,
computational complexity,
,
Cook, S. A.,
Core Logic, , , , , ,
, , , , ,
, , ,
Cornman, James W., ,
data type, x, , , , ,
, ,
Davidson, Donald,
de Kleer, J., , , , ,
, ,
decision problem, , , ,
, ,
for classical propositional
logic,
for contraction,
for intuitionistic
propositional logic,
for minimal propositional
logic,
for relevant propositional
logics,
for theoremhood,
for truth-table satisfiability,
for LR.,
deducibility,
defeater
rebutting,
undercutting, ,
degeneracy
of AGM-theory,
Degeneracy Theorem, ix,
dependency network, , ,
, , , , , ,
, , , , , , ,
, , ,
, , , ,
, , , ,
, , , , ,
, , , , ,
, , ,
, , ,
, ,
formally defined,
dependency networks
examples of,
formally defined,
deterministic polynomial
time,
directed acyclic graph, ,
Double-Negation
Elimination, ,
Double Twig,
downward amplification, ,
, , , ,
Downward pass, , ,
, , , ,
, , ,
, , , ,
, ,
downward sweep,
doxastic justifications,
doxastic status,
Doyle, J., , , , ,
,
Duhem, Pierre, ,
Duo, ,
dynamics
Newtonian,
of belief, , , , ,
of theory change, ,
Eiter, Thomas, xi,
entrenchment, , , ,
entrenchment, relative, , ,
, , , ,
entry (in a belief scheme),
, ,
evidential series, ,
excision set, , ,
, , ,
,
existence claim,
expansion, , , , ,
, , , , , ,
, , , , ,
Frege, Gottlob, ,
Friedman, Harvey M., xi, ,
, , , ,
Fuller, Timothy, xi
Fumerton, Richard,
Grdenfors, Peter, viii, ,
,
generality,
geometry of network, , ,
Gerhard, Gentzen,
Gottlob, Georg, xi, , ,
Henkin, Leon,
heuristics, ,
Hexagon, , , ,
Heyting, Arend,
Hilbert, David,
hitter, ,
perfect,
Hobsons choice, ,
Hume,
idealizing assumptions,
infinite regress,
infinitism,
informational economy,
criterion of,
intuitionist, , ,
intuitionist, cheeky,
intuitionistic relevant logic,
,
inviolable core of logic, ,
, , ,
item, justified in a belief
scheme,
Johannson, Ingebricht,
JTMS, , , ,
JTMS-theory, , ,
justificatory steps,
Kanigel, Robert,
kernel, , , , , ,
, ,
Kim, Joongol, xi
Klein, Peter, xi, ,
Kolmogorov axioms,
Komalatammal,
Kripke, Saul,
Kuhn, Thomas,
Kvanvig, Jonathan,
Law of Commutativity of
Addition, ,
Law of Excluded Middle, ,
Law of Non-Contradiction,
, ,
Left-rules,
Levi identity, ,
Levi, Isaac, xi, , , ,
, ,
Lewiss First Paradox, , ,
Lock
Black, , , , , , ,
, , ,
White, , , , ,
, , , , ,
logic
autoepistemic,
classical, , , , ,
classical first-order,
classical propositional, ,
core, , , , , , ,
, , , , ,
, ,
default,
intuitionistic, , ,
intuitionistic first-order,
minimal propositional,
non-monotonic,
relevant propositional,
Logical Closure Principle,
logical consequence,
logical falsehood,
logical reform, ,
logical truth,
logically minimal implying
set, , ,
MacPherson, James, xi
macro-level, , ,
Makinson, David, viii, ,
, ,
Marek, Viktor, xi, , ,
maxichoice contraction,
McDermott, D.,
melanism, ,
memory space, ,
methodological
requirements,
Meyer, Robert,
micro-level, ,
model,
modesty,
Moisa, Christina, xi
Moore, R. C.,
mutilating,
mutilation,
en bloc,
maximal,
minimal, , , , , ,
, , , , ,
, , , , ,
, , , ,
INDEX
343
Namagiri, ,
Narasimha, ,
natural deduction rules, ,
network, stratified,
Neurath, Otto,
neuropsychological reality,
,
Niedere, R., , ,
nodes, , , , , ,
, ,
black, ,
white, ,
non-monotonicity, ,
non-well-foundedness,
normal form (of proof), ,
NP-complete, , , , ,
, , , ,
, ,
NP-hard, ,
NP-oracle,
occluder, , , , ,
,
occluder, example,
Ogden, William, xi
optimization,
Osherson, Daniel, xi
outlist,
P=NP problem,
PA,
Pappas, George, xi
paradigm shifts,
paragon
logical,
paragon, logical, , , ,
, , ,
Parikh, Rohit, ix, xi
partial meet contraction
function,
Pascal, Blaise, xii
Peano Arithmetic,
Pearl, Judea, xi,
pedigree, justificatory or
epistemic, , , , ,
, , , ,
pluralist,
Poincar, Henri,
Pollock, John,
344
INDEX
polynomial hierarchy, , ,
, , ,
preferences, sacrificial,
Principle of Avoiding
Arbitrariness,
Principle of Avoiding
Circularity,
principle of positive
undermining,
Prolog, x, , , , , ,
, , , ,
propositional justifications,
PSPACE-complete,
pyramid, ,
quantifiers,
quietist,
Quine, Willard Van Orman,
vii, , , , , ,
, ,
QuineDuhem Problem, ,
Raffman, Diana, xi
Rajagopolan, T. K.,
Ramunajan, ,
Ramunajan*, , ,
rationality maxim,
rationality of belief revision,
Recovery, viii, , , , ,
, ,
reflexive stability, , , ,
,
refutability,
regress, justificatory, , ,
, ,
Reiter, R.,
relativity theory,
relevance logic R of Anderson
and Belnap,
relevantist, ,
Representation Theorem, ,
,
revision, , , , ,
algorithm,
logical,
Rhombus, , , ,
Right-rules,
Roche, William, xi
Russell, Bertrand,
Ryan, Mark, xi
safe contraction,
saint
computational, ,
logical, , , , , , , ,
Sartre, Jean-Paul,
satisficing, ,
saturatable contraction,
Scharp, Kevin, xi
Schumm, George, xi
search problem, , ,
,
Searle, John,
Segerberg, Krister, xi
selection function,
sequent rules,
Shapiro, Stewart, xi, ,
Simon, Herbert,
simplicity, ,
skepticism, ,
Skinner, Quentin, x
Smiley, Timothy, xi,
Smith, Ian, xi
Solo, ,
Sophies choices, doxastic,
Sosa, Ernest, , ,
soundness,
spectrum,
Sperner, Emanuel,
Spohn, Wolfgang, xi
Sprig, , ,
statics, , ,
Statman, R.,
step, , , , ,
initial, ,
transitional, , ,
steps, , , , , ,
,
informally explained,
initial v. transitional, ,
transitional,
subnetwork, , ,
, , , ,
, , , ,
, ,
T-closed,
Supowit, Ken, xi,
supplementary
AGM-postulates, ,
,
surrendering, , , ,
, ,
steps,
switching, , ,
syndrome
Chapter , vii
Firth of Forth Bridge, x
Tarski, Alfred,
Thagard, Paul, xi,
theoretical finitude,
thoroughgoing,
theory
in the logicians sense, ,
Thistlewaite, P.,
time, , , , , , ,
non-deterministic
polynomial, ,
polynomial,
TMS, , , ,
assumption-based,
transitivity of proof, ,
Truszczynski, M.,
truth tables,
Twig, , , ,
Ullian, Joseph, , , ,
,
undecidability
of first-order logic,
undecidability of R,
undecidability of first-order
logic,
unwanted,
Upward pass, , ,
, , , , ,
, , ,
Urquhart, A.,
Uzquiano-Cruz, Gabriel, xi
vacuous discharge,
virtues of scientific
hypotheses,
Washatka, John, xi
web of belief, , ,
Wedge, ,
well-foundedness,
Whitehead, Alfred North,
ZermeloFraenkel Set
Theory,
ZFC,
Zigzag, , , ,
INDEX
345