I

I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
362
A Hierarchical Approach to Designing
Approximate Reasoning-Based Controllers for
Dynamic Physical Systems
Hamid Berenji*, Yung-Yaw Chent, Chuen-Chien Leet, Jyh-Shing Jang S. Murugesan1
Artifcial Inteligence Research Branch, Mal Stop: 244-17
NASA Ames Research Center, Mofett Field, CA 94035
Abstract
This paper presents a new technique for the design
of approximate reaoning-baed controllers for dy
namic physical systems with interacting goas. In
this approach, goals are achieved baed on a hier
archy defned by a control knowledge bae and re
main highly interactive during the execution of the
control tak. The approach ha been implemented
in a rule-baed computer program which is used
in conjunction with a prototype hardware system
to solve the cart-pole balancing problem in rea
time. It provides a complementary approach to
the conventional analytical control methodology,
and is of substantia use where a precise mathe
matical model of the process being controlled is
not available.
Introduction and Motivation
Expert huma controllers ofen perform superbly un
der conditions of uncertainty and imprecision using
manly approximate reaoning. They select control ac
tions baed on a quick asessment of the process which
they are controllng. Control theorists have success
fully dealt with a lage clas of control problems by
mathematically modeling the process and solving these
analytical models to generate control actions. How
ever, the analytical models tend to become complex
and infeaible to use, especially for lage, intricate sys
tem. The non-linear behavior of may practical sys
tem maes the analytical approach even more dif
cult, sometimes impossible.
Starting with Mamdani and Assilian[Mamdani 75],
who baed their work on Zadeh's pioneering work on
fuzzy set theory [Zadeh 65], an alternative method in
sterling Federa Systems
Dept. of Electrica Engineering, Nationa Tawan Uni
versity, Tapei, Tawan, R.O.C.
toept. of Electrica Engineering ad Computer Scence,
University of California, Berkeley, 94720
iDept. of Electrica Engineering ad Computer Scence,
University of Caifornia, Berkeley, 94720
'ISRO Satelite Center, Banglore 560017, India
design of controllers ha been proposed. This tech
nique, generally known a fuzzy control, ha experi
enced much success, especially in recent years. These
controllers mimic the performance of human expert
operators by encoding their knowledge in terms of
linguistic control rules which may contain fuzzy la
bels (e.g., HOT, MEDIUM, SMALL). Among the
successful applications of this theory are the guid
ance control of subway trains in the city of Sendai
in Japan [Yaunobu 85] and cement kiln control
[Ostergaard 77]. A recent survey of this feld ha been
provided by [Lee 90b]. In general, these controllers
have been especially efective for systems with a single
goal. In this paper, we provide a new method for de
signing approximate reaoning-baed controllers which
can achieve conjunctive and interacting goals. We com
pare our method with the state feedback control, a
popular approach in modern digital control.
This comparative study is made using computer
simulation and a hardware implementation of a cart
pole balancing system which represents a typical non
linea system. This interesting problem ha served a
a bais for study by may connectionist works (e.g.,
[Widrow 87]) and control theorists (e.g., [Shaefer 66]).
Learning of the control process for pole balancing ha
been studied by Michie and Chambers [Michie 68], Sel
fridge, Sutton, and Barto [Selfridge 85], [Barto 83],
and by Lee [Lee 90a,Lee &Berenji 89]. In this learning
reseach, the objective ha been to write a program
which can learn to keep the pole balanced.
The organization of this paper is a follows. With a
brief overview of approximate reaoning-baed control,
we introduce a new method for designing controllers for
dynamic physical systems. We then apply this method
to the cart-pole balancing problem. The results of our
simulations and hardware tests are discussed next. Fi
nally, we contrat the performance of a controller baed
on our new approach with a conventional analytical
controller.
Approxmate Reasoning-Based
Controllers
A difculty in employing A technique in rea-time
control is how to hadle impreci&ion in the knowl
edge expressed by huma expert operators. Fuzzy
set theory provide a faility to expres the imprecise
knowledge by using lingui&tic
'
variabl3 [Zadeh 75]. We
have argued elsewhere about the importance of han
dling diferent type of uncertainty in AI system (e.g.,
[Berenji 88a], [Berenji 88b]).
The baic idea in fuzzy control centers around the
labeling process, in which the reaing of a sensr is
translated into a label a done by huma operators.
For eple, in the context of controlling a nuclear
reator [Bernard 88], a observed reactor period (i.e.,
the rate of rise of the power) might be claifed a too
3hort, 3hor, or negative. It is important to note that
the transition between the labels are continuous rather
than abrupt. This means that a reator's period of 90
seconds might be termd too 3hort to degree 0.2, 3hort
to degee 1.0, and negative to degee 0.0 [Bernard 88].
A simila concept is used in our experiment: an angu
la position of say 5 degree might be called Po3itive
to a degee of .8 ad Zer (i.e., a label used to describe
very small angles) to degree of 0.2. This idea of par
tial matching plays an important role in fuzzy control,
and is related to the concept of a membership function
ud in fuzzy set theory where the boundary of a set is
not shap and the degree of member3hip species how
strongly a element belongs to a set.
The knowledge bae of an approximate reaoning
baed controller is a collection of lingui&tic con
trol rle3 which are describd using lingui&tic
varabl3[Zadeh 75]. For example,
IF X is A and Y is B THEN Z is C
is a linguistic control rule where X and Y are sen
sor reaings from the plat and Z correspnds to the
output (i.e., the recommended action). A, B, and C
are linguistic values such a LARGE, POSITIVE, etc.
which are repreented by membership functions
(
usu
ally in triagular or trapezoidal form). When patic
ula values for X and Y ae sensed, then thes val
ues are matched aganst the membership functions of
A and B respctively. As a result of this matching,
the degee that eah precondition is satisfed will be
known. Since sensor reaings usually trigger svera
control rule at the same time, a confict re3olution
strategy is needed. A Maz-Min compo3itional rle of
inferenc, a explained below, is commonly used.
Assume that we have the following two rule:
Rule 1: IF X is At and Y is Bt THEN Z is Ct
Rule 2: IF X is A2 and Y is B2 THEN Z is C2
Now, if we have Zt and Yt a the sensor reaings for
fuzzy variable X and Y, then their trth value3 are
repreented by
A1 (zt)
and
B1 (Yt
)
respctively for
363
Rule 1, where JA1 repreents the membership function
for A1. Similarly for Rule 2, we have JA2
(zt
)
and
B, (yt) a the truth values of the preconditions. Then
the 3trength of Rule 1 can be calculated by:
VI
A1 (zt
) A B1 (yl).
Simlarly for Rule 2:
a
2

A
,(zt
) A s,(Yt
)
The efect of the strength of Rule 1 on its conclusion
is calculated by:
c (w) a1 A c1 (w),
and for Rule 2:
c

(w) a
2
A c,(w)
.
This means that a a result of reaing sensor values
Z and Yt1 Rule 1 is recommending a control action
with
c
( ) a its membership function ad Rule 2
is recomending a control action with l
c
; ( ) a its
membership function. The confict-resolution process
then produces
c(w) c(w)Vq(w) [at
Ac1(w)]v[a
2
Ac,(w)]
where
c(w)
is a pointwise membership function for
the combined conclusion of Rule 1 and Rule 2. The
A
and V operators in above are defned to be the min
and functions respectiveiy
[
Maudai 75]. The re
sult of this lat operation
(
J
c(w))
ha to be translated
( defzzifet to a single value. This necesary opera
tion produce a nonfuzzy control action that best rep
resents the membership function of an inferred fuzzy
control action. The Center Of Area (COA) method
(see [Lee 90b]) c be used here. Assuming a discrete
universe, we have
z
-
Ei=
t w
;
c(w
;
)
-
Ej=l
c(w
;
)
where T is the number of quantization levels of the
output.
Hierarchical Control and Conjunctive
Goal Achievement
In this section we develop a method for designing con
trollers which (a) obey a hierarchical process in fo
cusing attention on a particular goal at each time
instace, and (b) c achieve interacting goas si
multaeously. This discussion is related in may
ways to recent A planing reseach where integrated
planing-execution-control architectures are being ex
plored (e.g., [Drummond 89,Breina 90]). In this sec
tion, we present a brief discussion of our method in the
general context of approximate reaoning and in Sec
tion , we demonstrate the use of this technique in the
domain of cart-pole balancing. The method include
the following steps:
I
I
I
I
I
.I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
1. Let G = {g1, g2, gn} be the set of goas that sys
tem should ahieve and mantain. Notice that for
T = 1 (i.e., no interating goas), the problem be
come simpler and may be handled using the earlier
methods in fuzzy control (e.g., see [Mamdai 75]).
2. Let G = p
(
G) where p is a function which asigns
priorities among the goas. We asume that such a
function can be obtained in a particular domain. In
may control problems, it is poBible to specifcally
asign priorities to the goals. For example, in the
simple problem of balancing a pole on the palm of a
hand and alo moving the pole to a pre-determined
location, it is pO ible to do this by frst keeping
the pole a vertical a pOible and then gradually
moving to the desired location. Although these goas
are highly interative (i.e., a soon a we notice that
the pole is faling, we temporarily set aide the other
goal of moving to the desired location), we still can
asign priorities farly well.
3. Let U {u1, u2, ... , un} where u; is the set of input
control parameters related to achieving g;.
4. Let A = { a1, a2, ... ,an} where a; is the set of linguis
tic values used to describe the values of the input
control parameters in u;.
5. Let C = { c1, c2, , Cn} where c; is the set of linguis
tic values used to describe the values of the output
z.
6. Acquire the rule set R1 of approximate control rule
directly related to the highest priority goa. Thee
rules are in the general form of
IF u1 is a1 THEN Z is c1.
7. Fori= 2 ton, subsequently form the rule sets R.
The formt of the rule in these rule sets is simila to
the one in the previous step except that they include
apects of apprzmately achieting the pre'iou8 goal:
IF Yi-1 i8 apprzimately achieved and u; is a;
THEN Z is c;.
The approximate ahievement of a goa in step 7
of the above algorithm refers to holding the goa pa
rameters within smaller boundaries. The interactions
among the goal g; ad goal g;-1 are handled by forming
rules which include more preconditions in the left hand
side. For example, let U as ume that we have acquired
a set of rules R1 for keping a pole vertical in the palm
of a hand. In writing the second rule set R2 for moving
to a pre-spcifed location, apects of approximately
achieving g1 should be combined with control parame
ters for achieving g2. For example, a precondition such
a the pole i8 almo8t balanced can be added while writ
ing the rule for moving to a specifc location. A fuzzy
set operation known a Concentration [Zadeh 72] can
be used here to systematically obtain a more focused
membership functions for the parameters which repre
sent the achievement of previous goas. By defnition,
364
Concentration is a unay operation which, when ap
plied to a fuzzy set A, results in a fuzzy subset of A in
such a way that reduction in higher grades of member
ship i much leS than the reduction in lower grades of
membership. In other words, by concentrating a fuzzy
set, members with low grades of membership will have
even lower grades of memberships and hence the fuzzy
set becomes more concentrated. A common concentra
tion operator is to square the membership function:
JCON(A)(Y) J(y)
ad a typical concentration operator is the term Very
which is also a Lingui8tic Hedge [Zadeh 72]. For ex
ample, the result of applying the operator Very on a
fuzzy label Small is a new precondition Very SmalL
The Cart-Pole balancing problem
In the cart-pole balancing problem, a pole is hinged
to a motor-driven cart which moves on rail tracks to
its right or its left. The pole ha only one degee of
freedom (rotation about the hinge point). The tak of
a controller in this system is to keep the pole balanced
within a certain small range of cart positions on the
rail.
We now apply the method developed in the lat sec
tion to this problem:
1. Identif the goal set:
G={position the cart at the bcaticn @ L th: trak,
kep the pole balanced}.
2. Assign goal priorities:
g1 = keep the pole balanced
g2= position the cart at the location 1g on the track.
3. Identif the control parameters:
Four state variables are used to describe the system
status at eah time step, and one variable represents
the force applied to the cart. Thee are:
( angle of the pole with respect to
the vertical line
i angular velocity of pole
1 horizontal position of the cart on
the rail
i velocity of the cart
F amount of force applied to the cart
to move it toward the left
or the right.
We categorize these parameters a the following:
u1 = { (, i} is the set of parameters related to keep
ing the pole vertically balanced
u2 = { 1:} is the set of paameters related to hori
zontal position control.
4. Identif the linguistic values for each input control
parameter:
Three labels are used to linguistically defne the
value of the four state variables: Positive, Zero, and
Negative. Figure l(a) illustrates the memberships of
365
NL NH NI Z F$ M f
Nemw Zm Fsmve
Y:

0
Figure 1: (a)- Three quaitative levels for f, i, Z and 2, (b)- Seven qualitative levels for F
these linguistic term. Hence,
0 ={Positive, Zero, Negative}.
a2 ={Positive, Zero, Negative}.
5. Identif the linguistic values for the output:
For force F, we W seven fuzzy labels: Negative
Smal, Negative-Medium, Negative-Lage, Zero,
Positive-Small, Positive-Medium, and Positive
Large. Figure 1(b) illustrates the membership func
tions aoiated with these labels. Hence,
C ={Positive-Large, Positive-Medium,
Negative-Lage}
c2 ={Positive-Large, Positive-Medium,
Negative-Large}.
6. Acquire rules for highest priority goa:
Nine control rule are usd to keep the pole vertically
baaced. Thee rule ae lited in Appendix A. A
eple is:
IF ( is Positive and i is Zero THEN F is
Positive-Sml .
7. Form lower priority rule sets:
Approximately achieving the frst goa in this prob
lem refers to keeping the pole almost balanced. A
suming this ha been ahieved, we form the rule set
for achieving a new cart position. For Cple,
IF the pole is almo&t balanced and the cart is on
the right side of the desired location Zg
THEN push the cart to the right
2

The pole is asumed to be almod balanced if ( is
Ver Small and i is Ver SmalL As defned earlier,
Ver Small C be as umed to be the Concentrated
form of the label SmalL However, for simplicity, we
use a triangular membership function with a much
smaller bae. This process results in more complex
rules with 4 preconditions:
IF ( is Very Small and i is Ver Small and 1 is
Po&itive and z is Po&itive,
THEN F is Po&itite-Medium.

Although the quatative labls usd ae the same, df-
ferent scaes may be used for eah control paraeters.
At frst, ths might M to contradct our intuition to
push the cat to the rght when it u aready on the right
side of the desired poition KQ+ However, we push the cat
to the right hard enough that the pole stats to to
the lef. The subsequent attempt to keep the pole balnced
W move the cat towad the center.
Four rules of the above form are used to perform
cart position control. These rules are also listed in
the Appendix A.
POLE is a rule-baed program written in the C la
guage and serves a the controller in this problem. It
consists of only 13 rules, nine of which are used to
control the angular position ad the others are used
to control the position of the cart. The formt of the
rules is simple, each one having two or four precon
ditions and one consequent. The man reaon for the
simplicity of these rules is that they are lingui&tic con
trol rle, and the term in the preconditions can cover
a lage clas of sensor reaings, each to a diferent de
gree.
A charateristic of the POLE program, well
sme other system baed on linguitic control, is that
at any instat of time, more tha one linguitic rule
ml!ht b tewy to fre. In this cae, POLE perform
coDct-relution using the heuristic Maz-Min rl of
compitin eplained in Section .
Simulations and Exeriments
I th etion, the performance of POLE is compaed
wth a State Fedback Controller (SFC). SFC is one
of the moder control technique which uses a control
law U = -b. U is the input variable of the physical
system, which i a rea number in single-input system;
is the state vaiable which i an nelement colum
vetor; k is the n-element row vector of feedbak gains.
The SFC formulation is baed on the state space repre
sentation of the controlled system. The equations gov
erning the cart-pole system are given in the Appendix
B3.
Simulation-Based Comparison: We frst tested
the performance of thes controllers using computer
simulation. A set of 7 poles of diferent lengths and
weights were used. The length of these poles varied
between 0.5 ad 2 meters and their weights varied be
tween 0.05 and 2.0 Kilogram. We use the notation
(Pole-#, Length(m), Weight(Kg)). The poles had the
following charateristics: (Pole-1, 1.0, .1), (Pole-2, .5,
.05), (Pole-3, 1.0, .05), (Pole-4, .5, .025), (Pole-5, 1.0,
.5), (Pole-6, 1.0, 1.0), and (Pole-7, 1.0, 2.0).
3Due to space limitations, we avoid descrbing the
lengthy process o modeling and system identifcation whic
wa required to deign the SFC controller
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I

o
C

I
0
0
I
I
I
I

I
<
G
M

<
o
I
I
I
I
I
366
Table 1: Comparison of the fuzzy controller (FC) and the State Feedback Controller (SFC)
Pole-1
b L
a. 0 overshot (dereesJ .33 1.00
Ma. undershot (degrees) .87 2.29
aetiling time (seconds) 3.5 4.2
Ma. K overshot (em) 14.7 17.1
a. K undershot ( Cm) .8 1.7
K aetil time (snds) 38.2 4.6
fU L@M 0mm W. P
(PHe Agw C)

Pole-2
bL
.34 1.11
.73 2.41
3.0 4.8
8.8 16.2
.5 2.9
41.9 6.5
Pole6
L b L
.25 .63
.38 2.52
5.3 3.00
19.5 21.6
1.6 0.0
45.9 2.8
PU LQM RtMW W-
(Cn Pmm 0Q

W
4
Z

Z
M
.J .1
*fWW
** M
4
e
Z

B
D
E

O
&

U
Z
~
J
ic
c
t
8 t
t
/ \

.' \

/
.

'
'

\
\
mP

l
\
\
* M

.
Figure 2: Simulation data - Achievement of goals by the two controllers
fuZZy L0K C0htmll0f
t8t0 fm80K C0htf0ll0f
"
~

e
f
e

z
c <
-
c
M

0
:
0 <
c
b
Y

c O
Z
..
<
: : .1
o6: o i 00
1m0 (S.) 1m (S.)
Figure 3: Simulation data - Interaction between ihe two control goals
M
W
19
ic
c
e '

c
0
C

0
c
Y

O
In eah experiment, we compared the performance
of the fuzzy controller with the perfoOce of the
state feedbak controller. In eah cae, the fuzzy con
troller performed better, with les under- and over
shoot. However, it took more time for the fuzzy con
troller to reah stability, espcially for controlling the
position of the cart. Table 1 suizes this diference
for three of the poles, over a total samplng time of 50
seconds and a simulation time step of 5 mili-seconds.
Figure 2 presents a graphical display of the perfor
mace of the fuzzy controller (FC) and the state feed
back controller (SFC) in controlling the pole's angular
position (9), and the cart position on the trak (z).
The interactions between these two control goas are
shown in Figure 3 for the fuzzy logic controller and
the state feedbak controller.
Experiment-Based Comparison: We imple
mented both control schemes in a hardware system.
The hardware system included an IBM PC-AT, a DC
motor, 2 potentiometers for sensing the pole angle and
the cart position, ad a Data Acquisition and Control
Adapter. The cart-pole combination wa driven by the
DC motor through an aluminum chain which matched
the teeth of the driving pulleys (one pulley on the mo
tor shat). A samplng time of 20 mili-seconds wa
used.
Figres 4 and 5 illustrate the performance of the
hardware system under thes two type of control. In
our experiment, the fuzzy logic controller succeeded in
balancing the pole praticaly from the beginning of the
testing. Som tuning wa needed to control the cat
position at desired location on the traks. Not unex
pectedly, the SFC controller achieved very good perfor
mace aso. However, this wa acomplished through
a long system modeling ad identifcation process (a
precise mathematical model ha to be aquired before
the SFC controller could be implemented).
Frther Experiments: Severa other tests have
been performed using the prototype system. Figure
6(a) shows the interation between pole angular con
trol (goal) and cart position control (goal2) when the
pole wa tapped twice: once aer 15 seconds and once
again aer 35 seconds. Figure 6 shows the same inter
action aer the rail traks were tilted about 7 degree
afer 20 seconds and un-tilted afer 45 seconds.
We summaize the comparison of these approahes
baed on the following criteria:
Deign complexity: The design of a fuzzy controller
doe not reqwre a complete model of the process.
SFC is model-baed ad it requires a precis mathe
matical model of the proces. The complexity of the
model wa not a problem in our experiment, since
analytical models were reaily available for the cart
pole balacing problem. However, for a lage clas of
non-lnear control problem, this issue is signicat.
367
Controller modifcation: Compared to the state feed
bak controller, a fuzzy logic controller may involve
more parameters for fne-tuning. In our experiment,
fne-tuning the fuzzy logic controller seemed to be a
little more difcult.
Robustness: The design of the state feedbak con
troller does not include uncertainty in the system,
while a fuzzy logic controller, by modeling an op
erator's knowledge, ha a lager tolerance for vaia
tions in the process parameters. In our experiments
with poles of varying lengths and weights, the fuzzy
controller wa more robust than the analytical con
troller. In the cae of Pole 7 (i.e., the heaviest and
longest pole), the fuzzy controller balanced the pole,
albeit with difculty; the state feedbak controller,
however, faled to balance the pole.
Conclusions
We have presented a new method for designing ap
proximate reaonig-baed controllers. The hierarchi
ca nature of the method provide a better framework
for designers to focus their attention when writig the
control rules or fne-tuning them later for smoother
performance. We have used POLE and its prototype
hardware development to compare the performance of
our method with that of an analytical state feedbak
controller. POLE produced results very close to its
counterpart analytical cont.ol!er d in L M
POLE's results surpased them4 We believe that these
results are good indications of the versatility of our ap
proah in general a a complement to the conventional
controllers.
Under the a umptions outlined earlier, our method
should provide a better faility in designing approx
imte reaoning-baed controllers. Further tests need
to be done in applying this approach to other domns.
The results reprted in this paper have been encourag
ing enough to stat a lager scale project in applying
our method to the rendezvous and docking operation of
the Spae Shuttle with the Spae Station or a satellite.
Acknowledgements: May thanks to Peter Fied
lad, Mark Drummond, and Philip Laird for their valu
able comments on earlier drat of this paper.
Appendi A: Controller's Knowledge
Base
D Negative PL Positive Large
ZE Zero PM Positive Medium
PO Positive PS Positive Smal
Yb Very Smal NS Negative Smal
NL Negative Lae NM Neative Medium
Rules used for agula position control:
Rulel:I is PO AND i is PO THEN F is PL
Rule-2:1 is PO AND i is ZE THEN F i PM
A video, whc ilustrates the performace of the hard
we system, is avalable fom the author.
I
I
I
I
I
-I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
1
I

I
L
368
03
rvu=cm

0.J
0Z
01
0.I
0.1
!

1
l
w.I
w.J
0 '.,t'-t\ - ----- -l
.1
4.J
0
2.J
0
0J
0.I
04
0
|
w.1
.J
J
A

g
1
V
l
X

4.J
_
10
0 H W M W
TmI)
Tw1w
Figure 4: Hadware data - Pole agular control
C4t

40
FyLxCmW
U W Z J0
1
!

5
L

JJ
_

0J
0.Z
.1
2
J
-_
A

g 0 l$ 20 J I0 3
Tm1w1
Tm1w\
Figure 5: Hadwae data - Cart position control
"" '*Y

A
. j
.:
: `M : z: +:
1 W)
M
I
'

.

. . : z:
1|m m)
: t:
$0
ll
V
:: : : :
:
1A l
1
10
l

M

: : z:: : 9
1 m MI
ll
Figure 6: Hardware data - Interaction between pole angular control and cart position control After (a) tapping the
pole (b) tilting the rail tracks
Rule 3:1 6 is PO AND 8 is D THEN F is ZE
Rule 4:1 6 is ZE 8 is PO THEN F is PS
Rule 5:1 0 is ZE 8 is ZE THEN F is ZE
Rule- 6:1 6 is ZE AND 8 is D THN F is NS
Rule 7:1 6 is D AND 8 is PO THEN F is ZE
Rule8:1 6 is NE AND 8 is ZE THEN F is NM
Rule9:1 is NE AND 8 is NL THEN F is
Rules used for horizonta position conirol of the cat:
Rule 10:1 6 is VS AND 8 is VS x is PO AND z is
PO THEN F is PM
Rule-11 :I 6 is VS AND i is VS AND x is PO AND z is
ZE THEN F is PS
Rule-12:1 6 is VS AND i is VS AND x is NE AND z is
NE THN F is NM
Rule-13:1 6 is VS AND 8 is VS AND x is NE AND z is
ZE THEN F is NS
. Appenix B: Cat-pole balancing gover
mg equations
The governing equations of the ci-pole balancng prob
lem ae gven by the followg nonlnea diferentia equa
tions [Barto 83):
I+ ml[B2 sin ~6 co 6) legn( z)
me+m
where 0 is the agle of the pole with respect to ihi verica
line, K is the horizonta position of the cat, I is the drivng
force appled to the cat, g is the aceleration due to gravity,
me is the mas of the ct, m is the mas of the pole, l is
the haf-pole length, p is the coefcent of fiction of cat
on irak, ad lp is the coefcent of fiction of pole on ct.
[Bato 83)
[Bernad 88)
[Berenji 88a)
[Bereni 88b)
[Bresina 90)
References
Bato, A.G., Sutton, R. S., ad An
derson, C.W., Neuronlke Adaptive El
ements That Ca Solve Difcult Lean
ing Conirol Problems, IEE ba. on
Syst., Cybem., vol. SMC13, pp.
834-846, 1983.
Bernad, J. A., Use of a Rule Ba Sye
tem for Proces Control, IEE Conirol
System Magaine, October 1988.
Berenji, H. R., Responss io Saf
oti'sTeatet of Uncertainty in ,
Knowledge Enginerng Reviews, vol
ume 3, no 2, 1988.
Berenji, H. R., Treatment of Uncer
tanty in Artifc Intelgence, in: Ma
chne Intelligence ad Autonomy for
Aerospae Systems, . E. Beer ad
H. Lum, pp. 233-247, , 1989.
Breina, J ., ad Drummond, M., Inte
grating Planig ad Reactions: A Pre
lnary Report, Proc. of the Staford
Spring Symposium (Sesion on Pla
ning), 1990.
[Drm ond 89]
[Lee 90a]
[Lee 90b)
[Lee &Berenji 89)
[Madani 75)
[ichie 68)
[Ostergad 77)
[Selfidge 85)
[Shaefr 66)
[Widrow 87)
[aunobu 85]
[Zadeh 72)
[Zaeh 75)
[Zaeh 65)
369
Drummond, M., Goal Ordering in Pa
tial y Ordered Pla, Proc. of IJCAI-89,
Detroit, N pp. 960-965.
Lee, C.C., Self-Leanng rule-baed con
troller employing approxmate rean
ing ad neura-net concepts, io app
in Int. Jouma of Itelgent System,
1990.
Lee, C. C., Fuzzy Logc in Control Sys
tem: Fuzzy Logc Controller I ad ,
to appa in: IEE Tas. on System,
Ma ad Cybernetic, vol. 20, no. 2,
1990.
Lee, C.C., Berenj, H.R., An Intell
gent Controller Baed On Approxmate
Reasoning And Reinforcement Le
ing, Proc. of IEE Int. Symposum on
Intelgent Control, Abay, D 1989.
Madan, E. H., ad , S., An
experiment in linguistic synthesis wth
a fuzzy logc controller, Internatona
Journa of Ma-Maine Studie, Vol
ume 7, no.1, pp. 1-13, 1975.
Michie, D., ad Chabers, R., Boxe:
An experiment in Adaptive Control, in
Mane Intellgence 2, E. Dale ad D.
Michie Eds., Olver ad Boyd, Edn
burgh, 1968.
Ostergart, J M., Fuzzy logc control of
a heat exchanger proces, in Fuzzy Au
tomata ad Decsion Process, North
Rolad, 1977.
Selfidge, L. G., Sutton R.S., ad
Barto A. G., Traning ad Tang in
Rbotic, in Proc. of the 9th Interna
tiona Joint Conference on Artifc I
telgence, pp. 670-672, 1985.
Shafer, J. F., ad Canon, R. H., Con
tinuous Linea Systems: On the Control
of Unstable Mechanica System, 1966.
Widrow, B., The Origna Adaptive
Neura Net Broom-Baacer, It. Symp.
Cicuits ad Systems, pp. 351-357, May
1987.
Yaunobu, S. ad Miyaoto, S., Auto
matic Tran Operation System by Pre
dctiv Fuzzy Control, Industra Appl
ction of Fuzzy Control, pp. 1-18, 1985.
Zadeh, L.A., A fuzzy-st-theoretic inter
pretaton of lingustic hedges, Journa of
Cybernetic, 2, 4-34, 1972.
Zadeh, L.A., The concept of a lingu
tic vaable ad its application to ap
proxmate reaoning, Inform. Sc., Pai
1: vol. 8, pp. 199-249; Pai 2: vol. 8, pp.
301-357; Pari 3: vol. 9, pp. 43-80, 1975.
Zadeh, L.A., Fuzzy Sets, Information
ad Conirol, pp. 338-353, 1965.
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I