Anda di halaman 1dari 23

EUROPEAN

JOURNAL
OF OPERATIONAL
RESEARCH
ELSEVIER European Journal of Operational Research 103 (1997) 350-372

Supporting a complex audit judgment task:


An expert network approach
J e f f e r s o n T . D a v i s a, A n n e P . M a s s e y b,,, Ronald E.R. Lovell II a
a Department of Accounting, Management Information Systems and Marketing. Clarkson University, Potsdam, NY 13699-5795, USA
b Department of Accounting and lnfi)rmation Systems, School of Business, Indiana University, 1309E. lOth Street,
Bloomington, IN 47405, USA

Abstract

An auditor considers a tremendous amount of data when assessing the risk that the internal control (IC) structure of an
entity will fail to prevent or detect significant misstatements in financial statements. The myriad of relationships between 1C
variables that must be identified, selected, and analyzed often makes assessing control risk a difficult task. While some
general procedures and guidelines are provided, audit standards dictate no specifically set procedures and rules for making a
preliminary control risk asscssmcnt (CRA). Rather, the proccdures and rules are left mostly to auditor judgment. This paper
considers the appropriateness of applying artificial intelligence (A1) techniques to support this audit judgment task. It details
the construction of a prototype expert network; an integration of an expert system (ES) and a neural network (NN). The
rules contained in the ES model basic CRA heuristics, thus allowing for efficient use of well-known control variable
relationships. The NN provides a way to recognize patterns in the large number of control variable inter-relationships that
even experienced auditors cannot express as a logical set of specific rules. The NN was trained using actual case decisions of
practicing auditors. © 1997 Elsevier Science B.V.
Keywords: Expert network; Neural network; Expert system; Control risk assessment; Auditing

1. Introduction ber of variables in a problem increase and the speci-


ficity and measurability of their efficacy relation-
Many decision-making tasks do not lend them- ships diminishes, difficulties are exacerbated. The
selves to formulation through the sole use of quanti- literature suggests that builders of systems to support
tative models, nor simple intuitive problem solving complex decision-making should draw from and at-
(Rosenhead, 1992). Building and incorporating quali- tempt to combine a multitude of paradigms (for
tative and quantitative reasoning and modeling into a example, decision support, neural networks, case-
decision-aiding system is a challenge for practition- based reasoning, and models with rules or objects)
ers and researchers (Gupta, 1994; Silverman, 1995; (Silverman, 1995; Beulens and Van Nunen, 1988;
I.iberatore and Stylianou, 1993, 1995). As the num- Turban and Watkins, 1986; Yoon et al., 1993). This
paper analyses the process of building a prototype
intelligence-based system for application to a com-
plex problem within the field of auditing. The proto-
"Corresponding author. E-mail: amassey@juliet.ucs.indiana. type system is constructed as a decision-support tool
edu. for auditors analyzing the internal control (IC) struc-

0377-2217/97/$17.00 © 1997 Elsevier Science B.V. All rights reserved.


PII S0377-221 7(97)001 25-2
J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372 351

ture of a business entity in order to derive a prelimi- This paper, therefore, contributes to the literature
nary control risk assessment (CRA). It is designed as on the practical application of AI-based systems to
an expert network that combines two AI paradigms audit judgment tasks in at least two ways: (1) consid-
- expert systems (ESs) and neural networks (NNs). eration of the application of hybrid intelligent com-
A review of the literature suggests that AI tech- puting techniques to an area of auditing that is at
nologies such as ESs and NNs have found only present making almost no use of them, but where
limited use as decision aids in audit practice. More- there is developing interest and a potential for bene-
over, developers of such systems have generally fits to be accrued from their use; and (2) develop-
relied on singular means of knowledge representa- ment of an integrated ES and NN prototype system,
tion and reasoning. Intelligence-based systems in the CRA Expert Network, constructed for a PC-based
auditing have been primarily employed as traditional environment using tools that lead to a PC-based
knowledge-based systems (O'Leary and Watkins, product appropriate for field experimentation.
1991; Gray et al., 1991; Eining et al., 1994; Smith, The remainder of the paper is organized as fol-
1996), traditional statistical systems (Bell et al., 1990; lows. Section 2 describes the nature of the audit task,
Eining et al., 1994; Scott and Wallace, 1994), and as providing the background to the selection of an
stand-alone neural network systems (Bell et al., 1990; expert network approach. Section 2 also describes
Coakley and Brown, 1993). the principles behind the integration of ESs and NNs
This paper forms a part of a wider study into the and introduces the prototype CRA Expert Network.
increasing use of AI technologies in audit practice. Section 3 details the design and construction of the
As for any system, effectiveness and efficiency gains prototype system, including a description of an ex-
are not automatic or achievable without users accept- periment conducted with practicing auditors used to
ing the system as a decision aiding tool. Thus, obtain judgment data for training the NN. Section 4
system design, development and implementation are contains brief concluding comments.
key factors in promoting acceptance (Gillett et al.,
1995). Although design and development include,
for example, traditional system aspects such as the 2. O v e r v i e w of domain and system architecture
design of the user interface, etc., it may be that the
characteristics of the underlying system technologies When a public accounting firm audits a business
- here ESs and NNs - also have a major impact on entity, the potential exists that the auditors may not
user acceptance (Davis, 1994; Eining et al., 1994). discover material misstatements in the entity's finan-
More specifically, users such as auditors may be cial statements. The likelihood of not discovering a
leery of technology that appears to take the decision significant misstatement is called audit risk. During
out of their hands. However, acceptance and use may the audit process, the business entity's internal con-
be enhanced as users gain a better understanding of trol (IC) structure is evaluated to determine the
the characteristics, benefits and limitations of these nature, timing and extent of audit tests to be per-
intelligence-based systems, and the role of the tech- formed that will be most effective in reducing audit
nology as a decision aid as opposed to a decision risk. The purpose of the IC structure is to prevent
maker. a n d / o r detect erroneous, fraudulent, or missing ac-
Thus, a motivation for the development of the counting transactions. Consequently, a complete and
prototype system is to enable future research de- proper assessment of this structure is critical to a
signed to explore practitioners' acceptance (or rejcc- successful audit.
tion) of such systems. However, the objective of the The attempt to build intelligence-based systems to
research presented in this paper is narrower in scope. support audit judgment tasks, such as control risk
More specifically, this paper focuses on: (1)demon- assessment (CRA), is dependent on the acquisition
strating the appropriateness and applicability of an and representation of knowledge and heuristics
expert network within audit practice; and (2) describ- gained through audit experience (Deng, 1994). Gen-
ing the methodology employed to design and de- eral audit theory and heuristics may be captured and
velop the prototype system. represented using logical constructs. But, obtaining
352 J.T. Davis et al./ European Journal of Operational Research 103 (1997) 350-372

knowledge from auditors for making small yet im- single representation and reasoning approach (Lymer
portant distinctions between situations in the form of and Richards, 1995; Frisch and Cohn, 1991). ESs -
specific rules may be nearly impossible. That is, based on symbolic computation - are best at model-
assessing the IC structure potentially requires evalu- ing structured problem domains (or aspects thereof)
ating hundreds of variables with thousands of possi- that conform well to logical constructs. Conversely,
ble inter-relationships. It is not likely done by a NNs - based on numerical computation - can suc-
serial, step-by-step reasoning process, but rather by cessfully model problems that do not conform well
recognizing patterns in a given situation and reacting to explicit logical constructs. It is when these situa-
appropriately based on experience (Dreyfus and tions cross, such as in the CRA application, that the
Dreyfus, 1986; Anderson, 1983). Evaluating these combined capabilities of a blended solution should
many complex relationships is a difficult task even be considered (Caudill, 1991; Medsker and Bailey,
for the most experienced auditors - articulating them 1992; Medsker, 1994). By combining the deductive
is essentially impossible. However, neural network reasoning approach of an ES with the inductive
systems can be used to automate judgment tasks that approach of a NN, difficult and somewhat unstruc-
require this pattern recognition. tured tasks may be performed. Integration takes ad-
Thus, the innate complexities of the audit judg- vantage of the strengths of each type of system,
ment task suggest that an appropriate approach to while mitigating the inherent weaknesses of each
system design is to integrate sub-systems of AI when used alone. Let us briefly examine the subsys-
techniques each of which address distinct aspects of tems of an expert network.
domain knowledge and reasoning (Lymer and
Richards, 1995). Given the characteristics of the 2.1.1. Components of expert networks
CRA task, the prototype system is constructed as an ,sognitive models, focused on representations of
expert network reflecting the integration of an expert knowledge, suggest that human problem solvers or-
system (ES) and neural network (NN). Briefly, the ganize knowledge through the use of propositions. A
ES incorporates general audit theory and well-known proposition, in the form of an if-then rule, is the
control relationships using a logical set of explicit atomic building block of rule-based expert systems.
rules. The NN in the CRA Expert Network is used to Domain knowledge - generally elicited from a sin-
recognize and establish patterns among the large gle or multiple subject matter experts - is codified
number of inter-related variables inherent in the task. into a series of linearly executed discrete rules (Solso,
When exercised, the ES provides the user interface 1988; Greeno, 1973; Johnson-Laird, 1983; Minsky,
and evaluates the complexity and basic structure of 1986). However, significant development barriers
an entity's internal controls. If the ES determines arise because experts often cannot articulate knowl-
that the IC structure is sufficiently complex to war- edge and complex relationships as discrete rules.
rant further analysis, the data collected during inter- Only when asked do they produce a justification for
action with the ES is fed as input into the NN. The the judgments made. Even then, the justification is a
NN is stimulated by the input data and provides a generalization that more than likely has a fair num-
preliminary CRA. The NN evaluates approximately ber of exceptions (Deng, ! 994).
two hundred variables with thousands of potential Propositional logic is rarely a sufficient means to
inter-relationships used to make a preliminary CRA. represent complex reasoning (Kunz et al., 1987;
The following section examines the principles be- Jackson, 1990). Rather, ESs are a particularly good
hind expert networks such as the CRA Expert Net- approach for closed-system applications that have
work. literal and precise inputs that lead to logical outputs
(MacLennan, 1993). However, they are relatively
inflexible since performance degrades sharply when
2.1. Expert networks: An integrated approach they are applied to problems outside their original
scope (Jackson, 1990; Kunz et al., 1987; Edwards
An expert network is one form of integrated and Connell, 1989, p. 25). Moreover, changes in
systems designed to address limitations of using a domain knowledge structure and content often re-
J.T. Davis et al./ European Journal of Operational Research 103 (1997) 350-372 353

quire substantial system modifications to continue or A N N is a statistical modeling technique. How-


enhance system viability. ever, NNs can make less stringent assumptions con-
Conversely, NNs can analyze large numbers of ceming independence of variables and the shapes of
inter-related variables to establish patterns and char- underlying data distributions than other statistical
acteristics in situations where precise, discrete rules techniques, e.g., regression or multiple discriminant
are not known or discernible (MacLennan, 1993). analysis (Rumelhart and McClelland, 1986; Lipp-
Rather than depending on explicit knowledge as mann, 1987; Lacher et al., 1995). Rich discussions of
expressed by a domain expert, NNs model the im- the statistical aspects of NNs and their relation to
plicit relationships in exemplar domain data. The more traditional statistical models may be found in
continuous nature of the stimulus/response approach Ripley (1993, 1994), Cheng and Titterington (1994)
allows for efficient modeling of complex tasks. Sim- and Sarle (1994).
ply put, a neural network discovers its own numeric, While NN systems are flexible in terms of fault
continuous rules as it learns through the examples it tolerance to missing or noisy data, they do not have
is provided. NN systems may be able to perform some basic characteristics for flexible precise com-
certain types or parts of audit judgment tasks that are monsense reasoning, e.g., symbolic processing or
difficult and perhaps inappropriate for the capabili- interpretation of internally stored knowledge (Sun,
ties of other types of intelligent systems. 1994, p. 247). Unlike ESs, NNs have no inherent
An artificial NN consists of processing elements explanatory function 'module'. This has hindered
linked via weighted uni-directional signal channels acceptance in practice as it is not clear to a non-tech-
called connections to form a distributed, parallel nical user how the network derived a given conclu-
processing architecture (Rumelhart and McClelland, sion. However, research is being conducted to ex-
1986; Hecht-Nielsen, 1990). Each processing ele- tract comprehensible symbolic representations from
ment can possess local memory and carry out local- trained NNs (e.g., Craven and Shavlik, 1996).
ized information processing operations. The process-
ing element or neuron is the atomic building block of 2.1.2. Relationship o f the CRA expert network to
the NN (Fig. 1). NN paradigms differ in how the existing systems
processing elements are connected and the manner in A number of hybrid ES and NN systems have
which the weights are updated (Markham and Rags- been described in the literature. Kandel and Langholz
dale, 1995; Rumelhart and McClelland, 1986). (1992), Gallant (1993), Medsker (1995), and Sun

Xo= 1
(Bias)

Neuron or
XI
Processing
Element Neuron or
Inputs XL 1 Processing
Element
(Outputs from Output
other neurons
or processing
nodes)

Summation ] Transfer
Function ! Function

e.g. Sigmoid
Fig. I. Neural network processingelement structure.
354 J.T. Davis et al./ European Journal of Operational Research 103 (1997) 350-372

and Bookman (1995) have compiled research con- the user interface and conducts preprocessing of data
cerning the integration of symbolic and numeric that is fed into the NN. After the NN performs its
systems. These compilations include descriptions of pattern-matching activities, output is passed back to
expert networks applied to application domains such the ES for display. Advantages of loosely coupling
as natural language, signal processing, biology and the components of the prototype system are twofold:
medicine, management, and engineering (see particu- (1) system development is amenable to commercially
larly Medsker, 1995, pp. 3 9 - 5 6 ) . available software; t and (2) maintenance is reduced
The relationship between the ES and the NN in due to the simplicity o f the data file interface ap-
the CRA Expert Network is structured similarly to proach. The disadvantage of a great deal of redun-
systems reported by Lin and Hendler (1995), Macln- dancy that usually accompanies loosely coupled sys-
tyre et al. (1995), and Bataineh et al. (1996). Lin and tem has been largely avoided in the C R A Expert
Hendler (1995) use a NN to classify ballistic signals Network, perhaps due the nature of the problem
and the output is passed on to the ES for further more than any other reason. Before proceeding to an
processing and interpretation. Using the same neural in-depth discussion of the structure of the C R A
network software employed by the CRA Expert Net- Expert Network, the following two sections describe
work, Maclntyre et al. (1995) and Bataineh et al. the knowledge - and sources of that knowledge -
(1996) present expert networks for application within reflected in the ES and NN components, respec-
the electric utilities industry. The ESs in these appli- tively.
cations also provide processing inputs to the NNs.
Thus, while expert networks (and other forms o f 3.1. Expert system development
integrated systems) are not an explicitly new tech-
nique, we have found no evidence o f the technique The rules for the ES were primarily derived from
being applied to audit judgment tasks such as CRA. structured logic and questions found in Grant Thom-
However, given the characteristics of the task, there ton's internal control documentation and evaluation
is significant scope for consideration of their applica- software - Information and Control Understanding
bility to this domain. Furthermore, the ability to System (lnfocus) (Grant Thornton, 1992), which is
construct such systems using PC-based tools should used in audit and consulting practice. Infocus does
facilitate experimentation with them in the field. The not make any risk assessment itself and is not an AI
following sections detail the design and construction decision tool. However, it provides general relation-
of the CRA Expert Network. ships among IC variables to assist the auditor in
structuring the preliminary C R A process. The knowl-
edge and logic of Infocus were formalized in a
3. C o n t r o l r i s k a s s e s s m e n t (CRA) e x p e r t n e t w o r k logical rule structure. The interface provided by the
ES allows the auditor to document the IC data of a
As introduced earlier, the prototype CRA Expert given client. It should be noted, however, that the
Network was constructed by integrating an ES and user interface screens that are provided to an auditor
NN. The ES and NN components respectively ad-
dress two types of knowledge and logic relevant to
auditors: (1) well-known audit relationships - ap-
proximately 20% of the distinct internal control The ES component of the CRA Expert Network was devel-
oped using Microsoft Visual Basic 3.0. This tool provided the
structure variables - encoded as logical rules in the
means to: (a) develop an object-oriented GUI user interface; (b)
knowledge-base; and (2) more complex control t;ari- encode the well-known audit relationships in a rule structure that
able relationships determined by the 'trained' NN. logically drives data collection; and (c) directly link to a Microsoft
The CRA Expert Network uses a loosely coupled Access 2.0 database. The data collected as an auditor interfaces
model of an integrated system in which the CRA with the ES component is stored in an Access 2.0 database file. If
warranted, an ASCII text file of this collected data is fed to the
process is decomposed into separate ES and NN
NN as input. The NN, stimulated by these input variables, derives
components that communicate via data files (Medsker a CRA. The NN was created in NeuralWorks Professional I1/Plus
and Bailey, 1992; Medsker, 1994). The ES provides (NeuralWorks, 1992).
J.T. Daois et aL/ EuropeanJournal of OperationalResearch 103 (1997)350-372 355

during a given session are controlled by the logic duties. 2 However, the cases represented different
structure encoded in the ES - that is, data is neither situations in terms of firm size, complexity of control
requested of a user or screens provided that are not system, and strength of controls. The first case repre-
relevant to a given client. Use of Infocus' logic and sented an actual small-sized client, with low control
question was necessary because lnfocus was an inte- system complexity, yet a fairly strong IC structure.
gral part of the knowledge acquisition process con- This case was also compatible with a purely substan-
ducted with audit seniors - whose knowledge was tive audit approach. 3 The two cases from the audit
ultimately used to train the NN. In addition, lnfocus guide represented larger entities than the first case.
also provides a real world, accepted and tested basis The first, a closely held company, did not have as
for the majority of the ES logic. strong of an IC structure as the second, a larger
Added to the knowledge-base - that is not from public corporation that had a very strong IC struc-
Infocus' logic and questions - are CRA threshold ture. While both had computerized accounting sys-
rules and rules translating the preliminary CRA out- tems, the public corporation had a more complex and
put of the NN to a CRA category. The CRA thresh- well-controlled computer system.
old rules are used to control the current assessment Very broadly, while conducting a preliminary
of control risk while a session is in progress. For CRA, an auditor raises and reviews questions regard-
example, depending on the earliest data collected by ing the ICs (e.g., manual, computerized) of an entity
the ES component and the current threshold level that are designed to ensure that significant misstate-
setting, the ES may determine that further processing ments on financial statements will be prevented or
is not warranted, i.e., the data collection by the ES is detected. However, the questions that an auditor
halted and the NN is bypassed. This process is chooses to examine are left primarily to the auditor's
detailed in a following section. judgment. Thus, the purpose of working with the 64
auditors was to determine not only what their CRA
was for a specific case, but also what set of questions
3.2. Neural network development
each auditor was using during the CRA process.

3.2.1. Data source for training and testing 3.2.1.1. Collection of CRA process data. Using lnfo-
The data used to train the NN was from an cus, each auditor began their CRA by selecting
experiment conducted with 64 senior auditors from variables a n d / o r addressing specific questions from
Grant Thornton (see Davis, 1996). These subjects a set of potential judgment variables and questions
averaged 4.5 years of audit experience and bad related to the revenue cycle IC environment, EDP
completed a CRA an average of 37 times. environment, and accounting controls. Values as-
Each auditor was given one of three client cases. signed to selected variables and the responses to
The first represented an actual small sized client of chosen questions constituted each auditor's 'cue set'
the firm. The other two were based on example cases
found in the SAS No. 55 Audit Guide (AICPA,
1989). All three cases involved merchandising enti-
ties. The preliminary CRA was restricted to the sales 2 Segregation of duties refers to the principle that an individual
stream within the revenue cycle of each entity. The responsible for the conduct of a process cannot also be responsible
financial statement assertion group consisted of com- for the controls associated with that process. For example, if an
individual is responsible for receiving checks from customers,
pleteness, existence/occurrence, and valuation -
they should not be also responsiblefor recording those payments
core assertions related to the IC system goal of in the accountingsystem.
preventing and detecting errors. 3 Essentiallythere are two audit approaches. In the substantive
The three cases were chosen because the entities approach, the auditor elects to ignore the controls in place and
had computerized accounting systems and reflected focuses their analysis directly on the numbers represented in the
accounting statements. Conversely, in the control approach the
situations in which the auditor could potentially rely
auditor focuses first on an analysis of the IC structure of the client
on all three types of internal controls - manual in an effort to reduce the amount of substantive testing that will
controls, programmed controls, and segregation of follow.
356 J.T. Davis et al./European Journal of Operational Research 103 (1997) 350-372

- issues deemed as relevant to the specific case and, based on the numeric scale. The point estimate corre-
ultimately, assessment of control risk. The IC poten- sponds to the categorical judgment responses as fol-
tial variables/questions, and the possible values that lows:
each may be assigned, are presented in Appendix A.
The appendix reflects the complexities and relation-
Risk Category Point Estimate Interval
ships of the questions and variables that are poten-
tially considered by an auditor (including the 64
Limited (LTD) 0-25
auditors that participated in the experiment) during
Moderate (MOD) 2 6 - 50
the CRA task. Appendix A indicates the percentage
Slightly below 51-75
of the auditors, by case, that deemed each
maximum (SBM)
variable/question as relevant.
IC environmental features, EDP environmental Maximum (MAX) 76-100
features, accounting controls, and segregation of du-
ties operate as inter-dependent parts of the IC struc-
Each auditor's selected cue set and corresponding
ture. While these variables were included in this
derived CRA point estimate 4 make up one observa-
experiment and addressed by the auditors in various
tion. Although there were only three cases, the ob-
fashions, the complex inter-dependencies between
servations were distinct as each auditor indicated a
the variables makes it virtually impossible for an
different cue set for making their preliminary CRA.
auditor to explain the structure of their cue set, i.e.
For example, it is possible for an auditor to rely on
the relationships and strengths of relationships be-
programmed controls and choose a CRA of MOD.
tween items in the cue set, and the impact of this
Conversely, another auditor (examining the same
structure on the CRA. These inter-dependencies serve
case) could choose to rely on manual controls and
to illustrate the complexity of the domain and the
make the same preliminary CRA.
task facing an auditor when conducting a CRA. This
The 64 observations were used to develop and test
level of complexity also highlights the role that the
the NN model. The network training (within sample)
NN will play in structuring these relationships.
and testing (out-of-sample or hold-out) data sets each
In addition to identifying their respective cue set,
contained 32 observations. Once trained, using the
each auditor recorded: (1) whether s / h e planned to
32 within sample observations, the NN model serves
rely on manual controls, programmed controls,
as a proxy for the typical knowledge structure used
a n d / o r segregation of duties - indicating whether
by the experienced senior auditors in the training set.
more tests of controls as opposed to more substan-
The 32 observations in the hold-out sample were
tive tests would be used to reach an acceptable level
used to test the resulting NN model.
of audit risk; (2) the existing controls selected as key
controls, i.e., controls the auditor intended to rely on
3.2.2. Training and validation of NN
and test to determine if the control is operating as it
The CRA Expert Network employs a feedforward
should; and (3) any controls that s / h e felt were
classification NN that was trained using the back-
missing.
propagation learning algorithm (Rumelhart and Mc-
3.2.1.2. Collection of CRA output data. Following
analysis of the case, each auditor recorded their
preliminary CRA in Infocus. lnfocus provides a
choice of four risk categories: maximum; slightly 4 While it m a y seem tvdd to use a point estimate here, the
below maximum; moderate; and, limited. These risk motivation is rather simple. It allows for the use of a t-test to
categories are consistent with SAS No. 55 Audit c o m p a r e the N N ' s C R A (a value between 0 and 1) to that of the
auditors, rather than relying solely on classification a c c u r a c y . A
Guide. In addition to the categorical response, each
classification NN does not yield a c c u r a c y in relation to a t-test -
auditor recorded their CRA using a point estimate (0 just classification a c c u r a c y . Furthermore, this a p p r o a c h allows for
to 100 scale) which provides an indication of how the NN response to be m a p p e d to the C R A risk categories that are
close to the category border their judgment would be inherent in thc experimental task.
J.T. Davis et a l . / European Journal o f Operational Research 103 (1997) 350-372 357

Clelland, 1986; Tam and Kiang, 1992). A feedfor- with a smaller number of middle nodes were too
ward NN using sigmoidal activation functions is general and performed poorly overall. The final NN
mathematically capable of any continuous function architecture has 210 input nodes (control cues/varia-
and thus is applicable to a large variety of knowl- bles), 30 hidden (middle) layer nodes, and one out-
edge intensive tasks by distributing knowledge en- put node for the preliminary CRA using the auditors
coded into link weights that is learned from data point estimates within their chosen CRA category.
examples (Hornick et al., 1989). In the network the Delta rule summations were employed with sigmoid
processing element (refer to Fig. 1) takes inputs from transfer functions on the middle and output nodes,
other neurons and sums these inputs, X 0 . . . X,, us- while the standard root mean square error (RMSE)
ing a summation function, L i = EN_oWI,jXj. The function was used for "all network layers. Training
transfer function 'normalizes' the input summation stopped at 1600 iterations using the early stopping
by vectoring it into a predetermined range. The technique to avoid overfitting of the training sample
activation level of the processing element is deter- - the RMSE of the sample began to increase instead
mined by an activation function: X i = 1/(1 + e-~'). of decrease at that point. In addition, the NN was
The backpropagation learning algorithm was used analyzed after training at 4800 iterations. While ac-
to find the functional relationship between the inputs curacy - both RMSE and category accuracy - on the
(judgment cues/variables) and the target outputs training sample was better than at 1600 iterations,
(the auditors preliminary CRAs). A variety of net- accuracy on the hold-out sample and overall accu-
work configurations were tested during the design racy on the total sample declined.
phase. Networks with a larger number of middle The final trained network (at 1600 iterations) had
nodes learned the training data sample quite well, but a Pearson correlation coefficient of 0.869, and a
performed poorly on the test data sample. Networks CRA category accuracy rate of 72% for the training

Network Accuracy - Training Data

"E 0.9
0.7

iii,iiiiiil
0.5 ¢ Desired Output I
'~ 0.3 .._~.... Network Output I
! ,t Difference J
0.1
o -0.1
O
o -0.3

Observations

Network Accuracy - Testing Data


1
0.8 ~ :~:~i~i~ii~

r_.4k_. Desired Output


,¢ [ ~ Netw ork Output
0.2 ;~!~:i~i~!~:.~?~:~i~:N:~! ~: i~.:~4i~: %::~ ! ~ ~i:!-~i!~!i~::~~-:: ~ ~......... !~:~:;
,¢ Difference
0

O
o

Obs e rvations
Fig. 2. Network validation.
358 J.T. Davis et al./European Journal o f Operational Research 103 (1997) 350-372

General ~. [ Compu,er L I General Control


Processing ~ _1 Questions
Environment ~ .... 1 Overview ~ ]Computer Controls~ Phase 1
Questions (CPO) r////~, | (Gce)

. . . . t ounting
A Controls ~.
Phase 2

211
Vmmbles

1 2 3 .... 210 Input Layer

Neural i Phase 3
1 2 .... 30 Middle Layer
, -A .... LJ
I
1 Output

Fig. 3. Expert network: Knowledge base structure with embedded neural network.

data set. Testing using the hold-out sample resulted lower than the auditor. Fig. 2 presents the point
in a Pearson correlation coefficient of 0.695, with a estimate network accuracy in relation to both the
category accuracy rate of 78%. 5 The risk category training and testing samples.
prediction errors for the test sample included: four A paired t-test (see Koopmans, 1987, p. 325) was
observations that were one category higher than the also
• run on the predicted and actual network output.
assessment by the auditor: one observation each that No statistical difference was found for the training
were one category lower than the auditor, two cate- sample (mean difference = - 0 . 0 0 0 3 , one-tail p -
gories higher than the auditor, and two categories value = 0.48) or the testing sample (mean difference
= 0.0239, one-tail p-value = 0.24).
In addition, two classification networks were de-
The correlation coefficient is a measure of relationship be- veloped - the first employed backpropagation and
tween paired observations in two data sets - here, the relationship the second a radial basis function - using only the
between the auditor's point estimate and the model's point esti-
C R A classification categories and not the auditors'
mate for each observation. Category accuracy rates were derived
by comparing the NN point estimate output - based on an
point estimates. The hold-out sample classification
auditor's input cue set - to the auditor's chosen risk category. For accuracy for both these models was approximately
comparison, error rates for NN models built for financial distress 49%. Conversely, as described above, the point esti-
applications (Bell et al., 1990; Tam and Kiang, 1992) were in the mates for the trained network in the CRA Expert
range of 10-23%. However, it is important to note that the models
Network were within the classification ranges chosen
in those studies were much simpler than the NN developed in this
study. In those studies there were less than 10 input variables with
by the auditors 78% of the time. Clearly, for this
two response levels for output. Predicting a four category response data, using a point estimate within C R A category
with 210 input nodes is a significantly more difficult problem. provided superior model precision.
J.T. Davis et al./ European Journal of Operational Research 103 (1997) 350-372 359

(a)
!
! Ask Control
Environment Questions

Yes PHASE1

No

Yes Set CRT = MAX

1
Set CRT >= Slightly
Below Maximum

--No--
I SetManu,,Cootro,---)
Neural Network
Inputs = 0 I
I

yes
I

Set CRT >= Limited CRA = Current


Threshold

I Yes

Computer Processing
Overview (CPO)

Set CPO & GCQ---~


Neural Network
Inputs = 0
(
)

Fig. 4. (a) Expert network processes flowchart; (b) Expert network flowchart.
360 J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372

(b)

General Controls
Questionnaire
(GCQ)
PHASE1
No

Set GCQ Neural


Network No
Inputs = 0

l
Rely on
-'--!
Rely on
Manual
Controls
(MC)
t
Y~--

Manual . S -- Cu ent i l
Controls
(MC)

I
Yes
I

Accounting
. ~ ~Accounting Controls Accounting Controls
Controls
_bJ Questionnaire: Questionnaire:
-I Set PC Questions = 0 PHASE2 Set MC Questions = 0 Questionnaire:
Ask MC Questions Ask PC Questions Ask MC & PC
Questions

_ _ _ T . . _ _ _ _

General Environment PHASE3


Variable Values

Fig. 4 (continued).

3.3. CRA Expert Network Design cessing overview, and general control questions
(computer-related);
The ES and trained NN constitute the CRA Expert • Phase 2. Accounting controls phase; and,
Network. The system employs a three phased ap- • Phase 3. The Neural Network phase.
proach to determine a preliminary CRA as follows:

• Phase I. Environment phase that encompasses Fig. 3 illustrates the relationships among these
the general environment questions, computer pro- three system phases. Fig. 4 provides a flowchart
J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372 361

overview of the main logic contained within the heuristics as rules in the ES should improve an
CRA Expert Network and its three phases. auditor's control variable selection consistency as
Phases 1 and 2 address the logical data collection well as inter-auditor consistency.
aspect of the CRA Expert Network by incorporating The questions are designed to collect information
and presenting to the user the judgment variables/ concerning the control environment and computer
questions found in Appendix A. As noted previously, environment (if any) in which the accounting con-
the data collected during these phases is stored in a trols, addressed in Phase 2, must operate. This infor-
database file, which is fed to the NN. The NN mation is used by the ES, to make (if possible) a
(during Phase 3) accepts and utilizes values for the preliminary CRA without proceeding to Phases 2
following (as detailed in Appendix A): (1) the Rev- and 3. In addition, the information is used to set a
enue Cycle Environment variables; (2) the Planned control risk threshold (CRT). The purpose of the
Reliance variables; (3) responses to the Computer CRT is to prevent a final CRA that is below a
Processing Overview (CPO) questions; (4) responses predetermined level, given that certain conditions
to the General Control Questions (GCO, and; (5) all exist in the client environment.
the Accounting Controls variables. Values accepted Broadly, during Phase 1 (see the flowchart in Fig.
by the NN are: a ' 1 ' for a Yes response to a question 4 and the specific questions found in Appendix A),
or the indication that an Accounting Control variable the auditor is asked if s / h e plans to rely on the
is a key control; a ' - 1 ' for a No response to a control environment. If not, i.e., the auditor believes
question or that a specific Accounting Control vari- that the control environment is not reliable or testing
able is missing, and; a ' 0 ' indicating that the ques- the controls is inefficient, the ES sets the CRA to
tion/variable is not applicable to the CRA. As Fig. 3 maximum (MAX) and the task is completed. This
indicates, there exist 210 values passed to the NN - indicates that the auditor will conduct a purely sub-
each constituting a separate node in the input layer stantive audit approach. Conversely, a yes response
of the NN. indicates that a control audit will be conducted and
However, it should be emphasized here that not the auditor is then asked if s / h e plans to rely on
all questions/variables are applicable to all cases. segregation of duties. If so, reducing the control risk
Data is collected in Phases 1 and 2 via user interface threshold (CRT) to slightly below maximum (SBM)
screens - the presentation of which is controlled by is justified; if not, the CRT is set to maximum
the logic structure encoded in the ES. This logic (MAX). In either case, the auditor is asked if they
structure is illustrated in Fig. 4. The ES does not plan to rely on manual controls (MC). If s / h e does,
continue along a data collection path that is not the CRT is reduced to limited (LTD), and the system
relevant. For example, if the answer by the auditor to is triggered to obtain information concerning MC
a particular CPO question is ' N o ' , then the related during Phase 2. If no reliance is planned, the CRT
GCQ questions are not applicable to that case. The remains unchanged and the MC questions are omit-
ES then bypasses the related GCQ questions - the ted from Phase 2. The control variables related to the
user never sees those screens - and assigns '0s' omitted questions will be input to the NN as '0s'
(which are passed to the NN), representing not appli- (i.e., not applicable).
cable, to those GCQ questions (see Fig. 4(b)). The The next set of questions concerns whether any
following sections present specific details concerning part of the accounting system is automated. If there
each phase. is no planned reliance on MC and none of the
accounting system is automated, the CRA is set to
3.3.1. Phase 1: Environment phase the current CRT and the task is complete. If there is
During the environment phase, information neces- automation, the Computer Processing Overview
sary for the ES to apply broad internal control audit (CPO) will be completed, thus gathering information
heuristics is collected. These heuristics should be about the general computer environment and applica-
well known to most auditors. However, the possibil- tion complexity. The responses to these questions
ity exists that one or more of these heuristics could determine thrce things: (1) the level of computer
be omitted or misapplied. Thus, including these environment risk; (2) the level of computer applica-
362 J.T. Davis et aL / European Journal of Operational Research 103 (1997) 350-372

tion complexity, and; (3) the relevant questions to be 3.3.3. Phase 3: Neural network
included in the General Controls Questionnaire Assuming the auditor chooses to rely on manual
(GCQ). At this point the auditor may decide not to accounting controls, programmed accounting con-
rely on p r o g r a m m e d controls. If this is the case, no trois, or both, the responses collected from Phases 1
GCQ questions are asked and all the GCQ inputs to and 2 are fed into the NN (Fig. 4). Each question
the NN will be set to 0. In the event that manual and from the ES is mapped to one of the 210 NN input
programmed controls are not relied on, the system nodes. Given this input, the NN derives a prelimi-
sets the C R A to the current value of the CRT and nary CRA - an evaluation among 0 and 1 - which is
exits. Conversely, if the auditor is relying solely on presented to the user by the CRA Expert Network. In
manual controls, the system sets the program control addition to presenting the numerical evaluation, the
NN input variables to 0, and proceeds to Phase 2 ES transforms the response to the CRA risk cate-
where it asks the manual controls questions. If the gory. For example, if the response was 0.52, the risk
auditor elects to rely on program controls, the GCQ category would be slightly below maximum (SBM),
questions are presented for completion. although it is close to the S B M / M O D category
Finally, after completing the General Controls border. The point estimate from the NN does provide
Questionnaire (GCQ), the auditor (usually consulting information to the auditor as to how close the NN
with the computer audit specialist) will be given response is to a risk category border. Whether this
another chance to reject reliance on programmed information ultimately contributes to an auditor's
controls. If the auditor decides that computer con- Judgment a n d / o r leads to better system acceptance
trois are too weak to support reliance on pro- are empirical questions.
grammed accounting controls, all the GCQ inputs to
the NN will be set to 0. The system then moves on to 3.4. Directions f o r future development and research
Phase 2 for the manual accounting controls informa-
tion, if the auditor has previously indicated reliance The prototype system offers the opportunity for
on manual accounting controls. Once again, in the experimentation with practicing auditors, and pro-
event that manual and programmed controls are not vides a basis from which standards may be estab-
relied on, the system uses the current CRT to com- lished for a task in which no specifically set proce-
plete the CRA task. dures and rules are currently delineated. Future de-
velopments will center on addressing current limita-
3.3.2. Phase 2: Accounting controls" phase tions in the prototype CRA Expert Network.
The accounting controls phase includes the com- The first limitation relates the knowledge acquired
pletion of the manual accounting control questions, and represented in the system. More specifically, the
the programmed control questions, or both depend- system is based on a limited number of cases and the
ing on the decision path taken during Phase 1. If fact that the auditor subjects were from the same
either the manual controls or the programmed were firm. However, a sizable set of cases and additional
considered not applicable in Phase 1, the ES sets the auditors could be obtained from audits already com-
corresponding NN input values to 0 and does not pleted a n d / o r by providing auditors with additional
present the user with any questions related to these case scenarios.
controls. If either or both are considered relevant, the A related issue is whether the NN model using
appropriate Accounting Controls Questionnaire is point estimates within categories provides a more
presented to the user. precise decision-aid than a pure classification model.
The questions for Phase 2 relate to potential For this data, the extra information as to how close
accounting controls (see Appendix A). The auditor to a category border the auditors' judgments were,
responds to each question by labeling each individ- seemed to give the point estimate model an advan-
ual potential control as a key control (coded as a ' 1 ' tage over the pure classification models. Of course,
for the neural network); missing control ( - l); or not using a point estimate within category assumes more
applicable (0). Upon completion, the CRA Expert judgment precision on the part of the auditors. There
Network proceeds to Phase 3. may be a trade-off between judgment precision and
J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372 363

generalizability between the point estimate within and derive a conclusion, it does not provide any
the categories approach and the purely classification justification for the conclusion. A major concem for
approach. This empirical question should be tested in the auditor is being able to document audit judg-
future research particularly with regard to acceptabil- ments in order to defend his/her judgment, if neces-
ity by auditors. sary. For example, if the auditor agrees or disagrees
Second, most auditors are not computer audit with the suggested CRA, they must be able to docu-
specialists. When the computer processing overview ment their reasoning. This dilemma is similar to the
(CPO) indicates a complex system, a computer audit dilemma faced by the medical profession for using
specialist should be consulted. The computer audit intelligent systems as decision aids. However, inter-
specialist assists the auditor in deciding whether the pretation of how a NN produced a particular conclu-
computer controls are adequate to support reliance sion is not a trivial task, given that the very purpose
on accounting programmed controls. Thus, efforts of using a NN is to model complex relationships that
are underway to capture and represent in the system are not well understood. Much research in this area
this specialized knowledge. is left to be done.
Third, the system is currently narrow in scope. In
addressing the risk assessment task, only the sales
stream within the revenue cycle is included. How-
ever, the sales stream is general to nearly all busi- 4. Conclusion
nesses and important in most all audit situations. In
order to increase the usefulness of the system, a set The creation of the CRA Expert Network provides
of integrated systems each designed to deal with a good example of the potential applicability of
particular parts of all the accounting cycles should be integrate d AI technologies to audit practice and
developed. demonstrates the appropriateness of the integrated
Additionally, the system does not include assign- technique to the specific audit judgment task of
ment of accounting control variables to particular assessing control risk. Potentially, intelligent systems
accounting processes or control objectives. The NN such as this can provide several benefits to audit
only considers whether a particular control is a key firms, including: preservation, replication, and distri-
control and should be relied on, is a missing control, bution of expertise; new insights into the decision
or is not applicable. However, the same control may process; decision-aiding support: consistency of de-
apply to more than one accounting process. The cisions; and, increased productivity (Borthick and
system does not take into consideration these com- West, 1987).
pensating controls - controls that are in operation In conclusion, the CRA Expert Network takes
later in the accounting process stream that may advantage of both the deductive approach of an ES
strengthen the overall effectiveness of the IC system. and the inductive approach of a NN to provide a
Further design efforts will involve including this unique decision aid designed to support and facilitate
assignment of controls to processes and accounting the process of conducting a CRA. The system is
objectives. Efforts in this vein may take the form of tailored to auditors requirements in terms of data
additional rules in the knowledge-base, more neural collection and analysis and allows an auditor to take
network variables (and, perhaps even a different NN a more considered and structured approach. The
architecture), and more than three levels or values system offers audit firms the opportunity to improve
that a particular control variable could be given. upon the accuracy of the CRA by capturing the
Currently, research efforts with regard to the CRA expertise of multiple auditors and multiple client
Expert Network involve two main areas: (1) estab- experiences. Finally, the nature of the system en-
lishment of experiments with auditors in the field, courages efficiency in the analysis process as it
and (2) design and development efforts focused on follows a logic-driven data collection path, request-
enhancements to the user interface and the construc- ing the user to only address questions that are logi-
tion of an explanation module. While the NN pro- cally required to make a CRA. The validation results
vides the capability to model complex relationships to date indicate that an expert network shows promise
364 J.T. Davis et al. / European Journal o f Operational Research 103 (1997) 350-372

for addressing the large complex judgment processes Dollar value of transactions: High, Medium, Low
inherent in control risk assessments. While intelli- (63%, 93%, 100%);
gent systems are domain and task specific, this ap- Whether part of the cycle is automated (100%,
proach is conceivably transferable to other large 93%, 91%);
complex judgment tasks. Whether management override is a concern (81%,
60%, 61%);
Whether client has an internal audit staff that
Acknowledgements works to prevent or detect material misstatements
in the financial statements (89%, 33%, 87%).
The authors wish to thank Dr. George W. Krull,
Jr., National Director of Professional Development at Planned reliance on type of controls.
Grant Thornton, whom was the driving force behind Whether reliance is planned on manual controls
the experiment presented as part of this research. The
(required)
experiment took place in conjunction with the firm's Whether reliance is planned on programmed con-
national professional education program. Dr. Kmll trois (required)
also provided valuable comments with regard to the Whether reliance is planned on segregation of
experiment. In addition, we appreciate the direction duties (required)
provided by Stephen Yates, Partner and National
Director of Advanced Audit Techniques, who pro-
vided the cases, instruction and opportunity to utilize A.2. Computer processing overview (CPO) and gen-
Infocus for the study. eral controls questionnaire (GCQ) company wide
(Phase I)

Appendix A. List of judgment variables/ CPO question: Is a third party service organi-
questions 6 zation used to process all transactions involving
electronic data processing? (n/a, n/a, n/a)
A.1. General environment questions (Phase 1) Has the service auditor reported on the processing
of transactions? (n/a, n / a , n / a )
Revenue cycle environment variables. Does the report only cover policies and procedures
Volume of transactions: High, Medium, Low in place? (n/a, n / a , n / a )
(59%, 100%, 100%); Does the report test the policies and procedures?
(n/a, n / a , n / a )
Do we have a copy of the report? (n/a, n / a , n / a )

6 Each of the following represent an input node to the Neural CPO question: Are the client's computers in a
Network. There are 210 input nodes, including three each for dedicated physical area or facility? (n/a, 80%,
volume and dollar value (high, medium, low). These inputs can 87%)
take on values of 1 (Yes or Key Accounting Control), " - I' (No
or Missing Accounting Control), and "0' (Not Applicable to the GCQ objective: Computer system physical secu-
CRA).
A 'Yes' response to a particular CPO question will trigger a
rity is adequate.
corresponding GCQ analysis to determine the status of a GCQ Is entry to the computer and supervisor terminal
objective. areas restricted to those staff required for com-
The percentages in the parentheses, following each variable/ques- puter operations? (n/a, 80% 87%)
tion, represent the percentage of the auditors in the experiment Are lists of authorized personnel maintained and
that had included and addressed this item in their "cue set'. The
order respectively reflects values for the three experimental cases
kept up-to-data? (n/a, 80%, 78%)
- small client, nonpublic client, and public client. An n / a indi- Is a log of all visitors maintained and reviewed?
cates that the item was not applicable to the case. (n/a, 80%, 78%)
J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372 365

Do procedures exist to control access by non-DP used for production runs? (n/a, n/a, 96%)
personnel (e.g. engineers, janitors)? (n/a, 80%, Are application programmers prohibited from set-
87%) ting up and operating the computer, even during
Are DP management and security personnel noti- program testing? (n/a, n / a 96%),
fied when DP staff /eave the client's employ- Does management approve development at key
ment? (n/a, 80%, 87%) stages: feasibility systems proposals, outline sys-
Are employees who work outside normal opera- tems proposals, detailed system design including
tion hours properly authorized and adequately su- systems specifications, parallel running or system
pervised? (n/a, 80%, 87%) acceptance testing? (n/a, n/a, 65%)
Is the computer room physical security adequate
to restrict unauthorized access to program and data Computer Operations?
files? (n/a, 80% 87%) Is access to operators' consoles restricted to com-
puter operators and other authorized staff?. (n/a,
93%, 96%)
CPO question: Is there a separate EDP depart
Are computer operators restricted from performing
ment? (n/a, 93%, 96%)
programming functions or running unauthorized
GCQ objective: Separation of EDP duties is ade- jobs? (n/a, 93%, 96%)
quate. Are computer operators excluded from access to
Systems Programming? cash and accounting source documents? (n/a,
Are systems programmers prohibited from operat- 93%, 96%)
ing the computer system when production files or Is access to production source libraries by com-
application programs are resident? (n/a, n/a, puter operations' personnel restricted? (n/a, n/a,
91%) 96%)
Are systems programmers prohibited from chang- Are logs or records of computer system activity
ing production files and programs? (n/a, n/a, (jobs and program runs, reruns, abnormal termina-
96%) tions of jobs and programs, system console opera-
Are responsibilities for various processors or soft- tor commands) maintained and reviewed? (n/a,
ware products periodically rotated among mem- 93%, 96%)
bers of the systems programming staff?. (n/a, Are system log exceptions investigated and results
n/a, 70%) of the investigation documented? (n/a, 93%, 96%)
Are systems programmers' activities logged? (n/a, Are system utilities (ZAP, Super-ZAP, etc.) in use
n/a, 87%) and appropriately controlled? (n/a, 93%, 83%)
Are systems programmers' activity logs reviewed
by management? (n/a, n/a, 96%) Input/Output Scheduling?
Are system utilities (ZAP, Super-ZAP) in use and Is the preissuance of files for night and weekend
appropriately controlled? (n/a, n/a, 87%) shifts in accordance with approved job schedules?
Is there a written computer development strategy? (n/a, 93%, 87%)
(n/a, n/a, 52%) Is authorization required for unscheduled job re-
Are there adequate reporting procedures to moni- quests? (n/a, 93%, 87%)
tor the progress of computer development? (n/a, Are any scheduling overrides automatically logged
n/a, 39%) and subject to review? (n/a, 93%, 91%)
Are all terminals closed down at the end of opera-
Application Programming? tions? (n/a, 93%, 96%)
Is development work carried out by using separate Is there a secure location for sensitive output (e.g.
source libraries, separate object libraries, and sep- check printing)'? (n/a, 93%, 91%)
arate development machines? (n/a, n/a, 74%)
Are application programmers restricted from ac- Library maintenance?
cessing program libraries or data files which are Is formal authorization required to transfer pro-
366 J.T. Davis et aL / European Journal of Operational Research 103 (1997) 350-372

grams to the production library? (n/a, n / a , 78%) Are access and security features within the operat-
Is the transfer of files to the production library ing system utilized to restrict access to and
carried out by computer operations staff?. (n/a, amendment of program files? (n/a, n / a , 91%)
n / a , 78%) Is terminal activity automatically monitored and
Are all additions to the utility library authorized? rejected access attempts reported? (n/a, n / a , 91%)
(n/a, n / a , 83%) Does the access control software identify each
Do header labels hold reference and control data terminal via the logical address? (n/a, n / a , 91%)
(i.e. file name, generation numbers, volume ID, Does the access control software identify each
time/date stamps on creation, expiration dates, terminal via the physical address? (n/a, n / a , 91%)
control totals)? (n/a, n / a , 78%)
Does program library software, if any, monitor Is access control software used to restrict access to
and record program changes? (n/a, n / a , 91%) the data files?
Are master lists of authorized personnel main-
Are the activities of the EDP department (includ- tained, indicating restrictions on access? (19%,
ing management) appropriately segregated? 93%, 91%)
Is there an organizational plan that defines and Are access and security features within the operat-
allocates responsibilities and identifies lines of ing system utilized to restrict access to and
communication? (n/a, 93%, 70%) amendment of program files? (19%, 93%, 91%)
Are the responsibilities of the EDP department, Is terminal activity automatically monitored and
accounting department and other DP users clearly rejected access attempts reported? (19%, 93%,
defined? (n/a, 93%,96%) 91%)
Are there appropriate procedures for: rotation of Is there encryption of all sensitive data files?
duties/shifts, holiday arrangements, and termina- (15%, 93%, 91%)
tion of employment? (n/a, 93%, 91%) Does the access control software identify each
Are source documents handled within data pro- terminal via the logical address? (11%, 93%, 87%)
cessing only by data preparation and data control Does the access control software identify each
staff?. (n/a, 93%, 87%) terminal via the physical address? (11%, 93%,
87%)
Is a lack of segregation compensated for by in- For data base applications, is access to the data
creased management supervision? possible only through the DBMS? (15%, 93%,
Is there independent management review and ap- 91%)
proval of various data processing functions? (n/a,
n/a, n/a) Are reports generated by such software reviewed
Are there controls over certain privileged func- by management?
tions, such as requiring that all program library Is the record of jobs actual/y run, and console
updates are assigned to designated personnel? logs, reviewed by supervisory personnel? (15%,
(n/a, n / a , n / a ) 93%, 87%)
Is the computer usage summary produced and
CPO question: Can separate u ~ r s access the sys- reviewed? (15%, 93%. 87%)
tem concurrently? (19%, 93%, 91%)

GCQ objective: Access control over programs A.3. Computer processing overview (CPO) and gen-
and data files is adequate. eral controls questionnaire (GCQ) accounting cycle
Is access control software used to restrict access to specific (Phase 1)
the production programs?
Are master lists of authorized personnel main- Note: The next two questions determine whether
tained, indicating restrictions on access? (n/a, the environment for the particular accounting cycle
n / a , 91%) is/is not low risk. The next ten questions determine
J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372 367

whether the application complexity for this account- Are logs of all modifications to the source code
ing cycle is simple or advanced. maintained? ( n / a , n / a , 91%)
Is access to and use of the application software
CPO question: Are transactions for this cycle restricted? ( n / a , n / a , 87%)
processed by a third party service organization?
(n/a, n/a, n/a) Are source modifications authorized, tested, and
Has the service auditor reported on the processing documented?
of transactions? Does testing follow the predefined objectives of
Does the report only cover policies and procedures the test? ( n / a , n / a , 74%)
in place? Does testing use test data files instead of produc-
Does the report test the policies and procedures? tion files for its tests? ( n / a , n / a , 91%)
Do we have a copy of the report? Is a chronological record of all program amend-
ments maintained? ( n / a , n / a , 91%)
CPO question: Is this application PC-based? (n/a, Is the system documentation updated to reflect
n/a, n/a) amendments? ( n / a , n / a , 87%)
Is this application running on a standalone Are modifications made by third-party vendors to
PC? existing systems adequately recorded and con-
trolled to ensure that all procedures are appropri-
CPO question: Does the client have the source ately updated? (n/a, n / a , 91%)
code for the computer programs used in this Is the vendor software tested in the same way as
accounting cycle? (n/a, n/a, 100%) in-house developments? ( n / a , n / a , 83%)
Are emergency fixes to production programs fully
GCQ objective: Access control over source code is reported and independently reviewed by manage-
adequate. ment? (n/a, n / a , 96%)
Is access to source code controlled? Are the tests of the modified source code properly
ls the copying or renaming of sensitive programs supervised by management? ( n / a , n / a , 91%)
or utilities prevented/detected? ( n / a , n / a , 91%) Do the tests of the modified source code use
Are utilities, which have been made widely avail- complete computer systems to test the interaction
able (e.g. for reporting), restricted to readonly of different programs? ( n / a , n / a , 91%)
access? (n/a, n / a , 96%)
Does the operating system provide operation log- CPO question: Do multiple applications share the
ging which records: all use of a compiler and the same data bases (files)? (n/a, n/a, n/a)
targeted files, access and amendments to the pro-
gram libraries, attempts to copy program files GCQ objective: Data base application control is
to/from the production library? ( n / a , n / a , 91%) adequate.
If source code is stored off-line on tape or car- Is there an integrated dictionary system?
tridge, is the location physically secure? (n/a, Is the data base administrator responsible for the
n / a , 96%) development of the data dictionary? (n/a, n/a,
If source code is stored off-line on tape or car- n/a)
tridge, is the individual responsible an individual Are the data base changes requested by the data
other than an applications programmer, systems base administrator approved by some other author-
programmer or computer operator? ( n / a , n / a , ity? (n/a, n / a , n / a )
96%) Is the distribution of the data dictionary restricted?
(n/a, n / a , n / a )
Could or does the client make modifications to the
source code? Is there a data base administrator?
Is there a computer program that reports all Is the data base administrator responsible for con-
changes to each program and program library? trolling, developing and maintaining the data dic-
(n/a, n / a , 91%) tionary? (n/a, n / a , n / a )
368 J.T. Davis et al./ European Journal of Operational Research 103 (1997) 350-372

Does the data base administrator monitor data all elements of the distributed data system? ( n / a ,
base usage? ( n / a , n / a , n / a ) n/a, n/a)
Are procedures established which allow access to Are software controls used to prevent update inter-
the data base only through the DBMS software ference on central databases in distributed sys-
and prevent unauthorized access while the data tems? (n/a, n / a , n / a )
base is not under DBMS software control? ( n / a ,
n/a, n/a) CPO Question: Are telecommunications or net-
Does the data base administrator approved and log works used in this accounting cycle? (n/a, 100%,
all changes to the DBMS library? (n/a, n / a , n / a ) 100%)
Is access to the computer operation area by the
GCQ Objective: Control over telecommunications
data base administrator restricted or supervised?
a n d / o r networks is adequate.
(n/a, n/a, n/a)
Is physical access to the terminals controlled?
Are utility programs under control of the data base
Is the physical access of microcomputers or termi-
administrator? ( n / a , n / a ,
nals maintained by either equipment locks or con-
Does the data base administrator maintain and
trol over location? ( n / a , 100%, 100%)
review a log of programs run? ( n / a , n / a , n / a )
Is the use of terminals controlled by passwords?
CPO question: Are there real-time updates to Are terminals automatically locked out after failed
files when transactions are entered? ( n / a , n / a , sign-on? (n/a, 100%, 96%)
n/a) Are terminals automatically locked out if the ter-
minal is inactive for a specified period of time?
GCQ objective: Control over real-time updates to ( n / a , 100%, 96%)
files is adequate. Are sign-on and sign-off procedures specified and
Are control totals from the transaction history file then verified by the computer system? ( n / a , 100%,
reconciled to the totals from the updated files? 96%)
(n/a, n/a, n/a) Is access to all data files that are in the process of
If memo updating (A method in which the system being updated (i.e. 'live' files) controlled by pass-
issues memo transactions to temporarily update a word? (n/a, 100%, 100%)
copy of the file and then actual update is per- Do users refrain from using common passwords
formed in batches overnight) is used, is the up- (first name, spouse's name, birth date, etc.)? (n/a,
dated copy of the file matched against the master 100%, 100%)
file used during actual on-line processing? (n Do users refrain from displaying passwords exter-
/a n/a, n/a) nally (e.g. post-it notes) on the terminals? (n/a,
100%, 96%)
CPO question: Is the data processing function tbr Is the use of batch files to log onto the system
this accounting cycle decentralized (Distributed prohibited'? (n/a, 100%, 96%)
Proce~ing)? (n/a, n / a , n / a ) Are passwords protected and changed on a regular
basis?
GCQ objective: Control over distributed process-
Is the display of passwords prevented during log
ing applications is adequate.
on? (n/a, 100%, 100%)
Is a software log maintained for all transactions
Upon termination or resignation, are employees
including errors and retransmissions? (n/a, n / a ,
immediately denied any computer access? ( n / a ,
n/a)
100%, 100%)
Are data dictionary tools used to document and
Are passwords stored using a on-way encryption
monitor the responsibilities for various data items
algorithm? (n/a, 100%, 96%)
or elements among the distributed network'? ( n / a ,
n/a, n/a) Are the system user rights transaction or applica-
Arc transaction logs maintained and reviewed for tion specific?
J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372 369

Are system user rights established, monitored and Are exception reports investigated and reconciled?
changed by a system or network administrator? (19%, 93%, 96%)
(n/a, 100%, 91%)
Is the ability to change system user rights pro- CPO question: Is there a significant loss of visible
tected by a system administrator or supervisor audit trail in this accounting cycle? (n/a, n / a ,
password? ( n / a , 100%, 91%) n/a)
Is the system administrator or supervisor password GCQ Objective: Control over a 'paperless' audit
changed frequently? ( n / a , 100%, 91%) trail is adequate.
Has the system administrator or supervisor pass- For paperless or EDI transactions, are control
word been changed from default password set- totals such as batch totals, sequence numbers and
tings? ( n / a , 100%, 91%) 'line item counts' maintained which: track missing
Is the confidentiality of the system administrator transactions, validate completeness, ensure one-
or supervisor Password maintained? ( n / a , 100%, time-only receipt of a transaction, match func-
91%) tional acknowledgment reports, review exception
Is the system administrator or supervisor password reports to ensure corrections? (n/a, n / a , n / a )
encrypted? ( n / a , 100%, 91%)
CPO question: Have there been any hardware or
Is the system accessible by modem?
software malfunctions which resulted in a loss of
Is data encrypted during transmission and on the
data in this accounting cycle? (n/a, n / a , n / a )
files? ( n / a , n / a , n / a )
Are the phone numbers changed periodically? GCQ objective: Backup control is adequate.
(n/a, n / a , n / a ) Was data restored by use of backup?
Are the phone numbers listed on the modem or Were there restart facilities which enabled pro-
terminal? ( n / a , n / a , n / a ) cessing to continue from point of interruption?
Are there procedures to prevent line tapping, un- (n/a, n/a, n/a)
listed numbers and auto connection of dial-up Were there procedures which identified the pro-
lines? (n/a, n / a , n / a ) cessing stage reached at the time of malfunction?
Is password and transaction code documentation (n/a, n / a , n / a )
secured against unauthorized access? ( n / a , n / a , Were there controls to ensure that program li-
n/a) braries (most recent versions) and data files (most
recent backup or copy) were restored subsequent
CPO question: Does the software in this account- to the failure or malfunction? (n/a, n / a , n / a )
ing cycle generate transactions or pass informa-
Was data restored by manual re-entry?
tion to other cycles? (19%, 93%, 96%)
Were controls (batch totals, control totals, etc.)
GCQ objective: Control over computer generated used to ensure that the restoration of data was
transactions is adequate. complete and accurate? (n/a, n / a , n / a )
Are the methods used in the program to generate
Was data lost and not restored'?
the data and related control record appropriate?
Was the nature of the data insignificant? (n/a,
(19%, 93%, 96%)
n/a, n/a)
Is there an adequate check over the accuracy of
Was the affect upon the financial statements im-
the data generated (e.g. reasonableness check,
material? ( n / a , n / a , n / a )
manual review of data generated, etc.)? (19%,
93%, 96%) A.4. A c c o u n t i n g control variables (Phase 2)
Are the results of the check for accuracy reviewed
and approved? (19%, 93%, 96%) Balancing receivable subledger (1 9%, 20%, 43%)
Are controls such as computer sequence checks, Balancing run to run control totals (11%, 13%,
computer matching or batch totals used to ensure 57%)
the completeness of input a n d / o r updates? (19%, Balancing the G / L with the subledgers (52%,
93%, 96%) 47%, 17%)
370 J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372

Canceling original documents (11%, 27%, 13%) Non-interactive range check edit (0%, 0%, 0%)
Checking by computer for duplicate entries (11%, Non-interactive reasonableness edit (0%, 0%, 0%)
27%, 30%) Non-interactive prior data matching (0%, 0%, 0%)
Checking by computer the numerical sequence of Non-interactive check digit (0%, 0%, 0%)
a file (7%, 20%, 39%) Performance of analytical procedures and investi-
Checking for a third party signature (7%, 13%, gation of unusual items (0%, 13%, 17%)
9%)) Periodic reconciliation of books to physical (7%,
Checking manually the numerical sequence of a 13%, 4%)
document, journal or report (0%, 20%, 17%) Periodic revision of budgeted amounts based on
Checking one-to-one (26%, 7%, 9%) updated information (0%, 0%, 4%)
Comparing batch totals (4%, 7%, 43%)) Physical access controls (7%, 13%, 4%)
Comparison of budgeted amounts to actual Physical safeguards (7%, 7%, 4%)
amounts (0%, 7%, 9%) Reconciliation of manual totals to run totals (0%,
Comparison of cash receipts listing to deposit slips 7%, 17%)
(0%, 27%, 17%) Reconciliation of master file balance to control
Computer generation of transactions (0%, 0%, account (0%, 34%, 9%)
o%) Reconciliation of master file balance to run totals
Dual control over cash receipts (11%, 20%, 4%) (7%, 7%, 35%)
Electronic authorization (0%, 7%, 0%) Reperformance (0%, 7%, 0%)
Independent review of edit reports for data file Restrictive endorsement of checks received (0%,
changes (0%, 20%, 13%) 7%, 4%)
Interactive dependency edit (0%, 0%, 4%) Reviewing adjustment transactions (0%, 7%, 4%)
Interactive document reconciliation (0%, 0%, 4%) Reviewing internal signatures (7%, 20%, 9%)
Interactive edit controls (0%, 7%, 0%) Reviewing permanent data file exception reports
Interactive existence edit (0%, 0%, 0%) (0%, 0%, 0%)
Interactive feedback edit (0%, 0%, 0%) Reviewing reference file data (0%, 0%, 0%)
Interactive format edit (0%, 0%, 0%) Software access controls (0%, 33%, 9%)
Interactive key verification (0%, 7%, 0%) Use of a Iockbox (7%, 7%, 0%)
Interactive mathematical accuracy check (7%, Use of prerecorded input (OCR, MICR, OMR)
20%, 9%) (0%, 0%, 0%)
Interactive prior data matching (0%, 0%, 0%) Verifying mathematical accuracy (7%, 7%, 4%)
Interactive reasonableness edit(O% 13%, 4%) Other: Review, compare invoice adjusted for out
Interactive range check edit (0%, 0%, 0%) of stock items with the packing slip (11%, 14%,
Interactive check digit (0%, 0%, 4%) 0%)
Matching to a previously validated document Other: Driver gets signature on annotated sales
(37%, 40%, 30%) invoice. (19%, 0%, 0%)
Matching to a previously validated file (0%, 27%, Other: Cost of goods entry is calculated from sales
52%) invoices and calculations reviewed (4%, 0%, 0%)
Matching to an authorized list (26%, 47%, 22%) Other: Credit Check on invoice customer (0%,
Monthly review of bank reconciliations (0%, 13%, 0%, 0%)
4%) Other: Compares amounts written off to aged trial
Monthly review of receivable (7%, 13%, 13%) balance and customer detail (0%, 0%, 0%)
Non-interactive format edit (0%, 0%, 0%)
Non-interactive mathematical accuracy (0%, 0%,
o%) References
Non-interactive edit controls (0%, 0%, 0%) A I C P A , 1989. A m e r i c a n Institute o f Certified Public Accountants.
Non-interactive existence edit (0%,0%, 0%) Audit guide: Consideration of the internal control structure in
Non-interactive dependency edit (0%, 0%, 0%) a financial statement audit. N e w York, NY.
J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372 371

Anderson, J., 1983. A spreading activation theory of memory. with the Auditing Section of the American Accounting Associ-
Journal of Verbal Learning and Verbal Behavior, 261-295. ation, New York.
Bataineh, S., AI-Anbuky, A., Al-Aqtash, S., 1996. An expert Grant Thornton, 1992. Information and control understanding
system for unit commitment and power demand prediction system version 2.08.
using fuzzy logic and neural networks. Expert Systems, 29-40. Gray, G., McKee, T., Mock, T., 1991. The future impact of expert
Bell, T., Ribar, S., Verchio, J., 1990. Neural nets vs. logistic systems and decision support systems on auditing. Advances
regression: A comparison of each model's ability to predict in Accounting 9, 249-273.
commercial bank failures. KPMG Peat Marwick Working Greeno, J.G. 1973. The structure of memory and the process of
paper, May 1990. solving problems. In: Solso, R.L. (Ed.), Contemporary Issues
Beulens, A., Van Nunen, J., 1988. The use of expert system in Cognitive Psychology: The Loyola Symposium. Wiley,
technology in DSS. Decision Support Systems 4, 421-431. New York, NY.
Borthick, A.F., West, O., 1987. Expert systems - A new tool for Gupta, U., 1994. How case-based reasoning solves new problems.
the professional. Accounting Horizons 1, 9-16. Interfaces 24 (6), 110-119.
Brown, C., Coakley, J., Phillips, M.E., 1995. Neural networks Hecht-Nielsen, R., 1990. Neurocomputing. Addison-Wesley,
enter the world of management accounting. Management Ac- Reading, MA.
counting, 51-57. Hedberg, S., 1995. Where's AI hiding? AI Expert, 17-20.
Brown, C., Phillips, M., 1990. Expert systems for management Hornick, K., Stinchcombe, M., White, M., 1989. Multilayer feed-
accountants. Management Accounting, 18-22. forwared networks are universal approximators. Neural Net-
Brown, C., O'Leary, D., 1994. Introduction to Artificial Intelli- works 2, 359-366.
gence and Expert Systems. Published Monograph, 1994. Jackson, P. 1990. Introduction to Expert Systems. Addison-Wes-
Caudill, M., 1991. CRA expert networks. Byte, 108-116. ley, Workingham, MA.
Cheng, B., Titterington, D., 1994. Neural networks: A review Johnson-Laird, P.N., 1983. Mental Models: Toward a Cognitive
from a statistical perspective. Statistical Science 9, 2-54. Science of Language, Inference, and Consciousness. Harvard
Coakley, J., Brown., C. 1993. Artificial neural networks applied to University Press, Cambridge, MA.
ratio analysis in the analytical review process. International Kandel, A., Langholz, G., 1992. Hybrid Architectures for Intelli-
Journal of Intelligent Systems in Accounting, Finance and gent Systems. CRC Press, Boca Raton, FL.
Management, 19-39. Koopmans, L. 1987. Introduction to Contemporary Statistical
Craven, M., Shavlik, J., 1996. Extracting tree-structured represen- Methods. Duxbury Press, Boston, MA.
tations of trained networks. In: Advances in Neural Informa- Kunz. J., Kehler, T., Williams, M., 1987. Applications develop-
tion Processing Systems 8. MIT Press, Cambridge, MA. ment using a hybrid A1 development system. In: Hawley, R.
Davis, E., 1994. Effects of decision aid type on auditors' going (I-M.), Artificial Intelligence in Programming Environments.
concern evaluation. Audit Judgment Symposium, Co-spon- Ellis Horwood, Chichester.
sored by Grant Thornton and University of Southem Califor- Lacher, R., Coats, P., Shanker, C., Franklin, L., 1995. A neural
nia, pp. 21-22. network for classifying the financial health of a firm. Euro-
Davis, J. 1996. Experience and auditors' selection of relevant pean Journal of Operational Research 85 (1), 53-65.
information for preliminary control risk assessment. Auditing: Liberatore, M., Stylianou, A., 1993. The development manager's
Journal of Practice and Theory, 16-37. advisory system: A knowledgebased DSS tool for project
Dcng, P., 1994. Automating knowledge acquisition and refine- assessment. Decision Sciences 24 (5), 953-976.
ment for decision support: A connectionist inductive inference Liberatore, M., Stytianou, A., 1995. Expert support systems for
model. Decision Sciences 24 (2), 371-393. new product development decision making: A modeling
Dreyfus, H., Dreyfus, S., 1986. Mind Over Machine: The Power framework and applications. Management Science 41 (8),
of llurnan Intuition and Expertise in the Era of the Computer. 1296-1316.
Free Press, New York, NY. Lin, C., Hendler, J., 1995. Examining a hybrid connections/sym-
Eining, M., Jones, D. and Loebbecke, J., 1994. An experimental bolic system for the analysis of ballistic signals. In Sun, R.,
examination of the impact of decision aids on the assessment Bookman, L. (Eds.), Computational Architectures Integrating
and evaluation of management fraud. Audit Judgment Sympo- Neuraland Symbolic Processes. Kluwer Academic Publishers,
sium, Co-sponsored by Grant Thornton and University of Norwell, MA, pp. 319-348.
Southern California, February 21-22. Lippmann, R., 1987. An introduction to computing with neural
Frisch, A., Cohn, A., 1991. Thoughts and after-thoughts on the nets. IEEE ASSP Magazine, 4-22.
1988 Workshop on Principles of Hybrid Reasoning. A1 Maga t.ymer, A., Richards, K., 1995. A hybrid-based expert system for
zinc: Special Issue, January 1991, 77-.83. personal pension planning in the UK. Intelligent Systems in
Gallant, S. 1993. Neural Network Learning and Expert Systems. Accounting, Finance and Management 4, 71-88.
MIT Press, Cambridge, MA. Maclntyre, J., Smith, P., Harris, T., Industrial experience: The use
Gillett, P., Bamber, E., Mock, T., Trot.man, K., 1995. Audit of hybrid systems in the power industry. In: Medsker, L. (Ed.),
judgment. In: Bell, T., Wright, A. (Eds.), Auditing Practice, Hybrid Intelligent Systems. Kluwer Academic Publishers,
Research, and Education: A Productive Collaboration. Ameri- Norwell, MA, pp. 57-74.
can Institute of Certified Public Accountants in Cooperation MacLennan, B., 1993. Continuous symbol systems - The logic of
372 J.T. Davis et al. / European Journal of Operational Research 103 (1997) 350-372

connectionism. In: D.S. Levine, M. Aparicio IV (Eds.), Neural Rumelhart, D., McClelland, J., 1986. Parallel Distributed Process-
Networks for Knowledge Representation and Inference. ing: Volumes I and II. The MIT Press, Cambridge, MA, pp.
Lawrence Erlbaum, Hillsdale, NJ. 110-146.
Markham, I., Ragsdale, C., 1995. Combining neural networks and Sarle, W., 1994. Neural networks and statistical models. In:
statistical predictions to solve the classification problem in Proceedings of the Nineteenth Annual SAS Users Group
discriminant analysis. Decision Sciences 26 (2), 229-242. International Conference. SAS Institute, Cary NC, pp. 1538-
Medsker, L., 1994. Design and development of hybrid neural 1550.
network and expert systems. In: Proceedings of the IEEE Scott, D., Wallace, W., 1994. A second look at an old tool:
International Conference on Neural Networks III, Orlando, FL, Analytical procedures. The CPA Journal, 30-35.
pp. 1470-1474. Silverman, B., 1995. Knowledge-based systems and the decision
Medsker, L., Bailey, D., 1992. Models and guidelines for integrat- sciences. Interfaces 25 (6), 67-82.
ing expert systems and neural networks. In: Kandel, A., Solso, R., 1988. Cognitive Psychology, 2nd edition. Allyn and
Langholz, G. (Eds.), Hybrid Architectures for Intelligent Sys- Bacon, Boston, MA.
tems. CRC Press, Boca Raton, FL, pp. 153-171. Sun, R., Bookman, L., 1995. Computational Architectures Inte-
Medsker, L., 1995. Hybrid Intelligent Systems. Kluwer Academic grating Neural and Symbolic Processes. Kluwer Academic
Publishers, Norwell MA. Publishers, Norwell, MA.
Minsky, M., 1986. The Society of Mind. Simon and Schuster, Sun, R., 1994. A two-level hybrid architecture for structuring
New York, NY. knowledge for commonsense reasoning. In: Sun, R., Book-
NeuralWare: Technical Publications Group, 1992. Neural comput- man, L. (Eds.), Computational Architectures Integrating Neu-
ing - A technology handbook for professional ll/plus and ral and Symbolic Processes. Kluwer Academic Publishers,
neuralWorks explorer. Pittsburgh, PA. Norwell, MA, pp. 247-282.
O'Leary, D., Watkins, P., 1991. Review of expert systems in Sutton, S., Young, R., McKenzie, P., 1995. An analysis of
auditing. Expert Systems Review for Business and Account- potential legal liability incurred through audit expert systems.
ing, 3--22. Intelligent Systems in Accounting, Finance and Management
Ripley, B., 1994. Neural networks and related methods for classi- 4, 191-204.
fication. Journal of the Royal Statistical Society B 56 (3), Tam, K., Kiang, M., 1992. Managerial applications of neural
409-456. networks: The case of bank failure predictions. Management
Ripley, B., 1993. Statistical aspects of neural networks. In: Barn- Science 38 (7), 926-947.
dorff-Nielsen, O., Jensen, J., Kendall, S. (Eds.), Networks and Turban, E., Watkins, P.R., 1986. Integrating expert systems and
Chaos - Statistical and Probabilistic Aspects. Chapman and decision support systems. MIS Quarterly 10 (2), 121-136.
Hall, London, pp. 40-123. Yoon, Y., Guimaraes, T., Swales, G., 1993. Integrating artificial
Rosenhcad, J., 1992. Into the swamp: The analysis of social neural networks with rule-based expert systems. DSS Special
issues. Journal of Operational Research Society 43 (4), 293- Issue on Artificial Neural Networks.
305.

Anda mungkin juga menyukai