Case Match Reduction through the Integration of

advertisement
Case MatchReduction through the Integration of
Rule-based and Case-based Reasoning Procedures
Huei-Pi (Ruby) Chen, Larry J. Wilkinson
NewcastleBusinessSchool,University of Northumbriaat Newcastle
Newcastle UponTyne, NE18ST, England
hpchen@msl.showtower.com.tw;
l.j~wilkinson@unn.ac.uk
Abstract
Case-basedreasoning utilizes past experiences as a key data
resource for future problem solving and is considered an
innovativetechniquein the development
of Artificial Intelligence.
However, the large storage requirements needed for such
proceduresresult in a heavyload in the ease base and indeed,
problemsof efficiency arise in case retrieval. The proposed
methodaimsto improvecase retrieval throughreducingthe case
memoryload. The authors propose a cognitive notion as a
framework for reducing case match. Using a cognition
framework,objective knowledge
is representedin the rule-based
part of the systemsand subjective knowledgein the case-based
reasoningpart. In this way,the comparisonsin case matchcan
be significantly reduced. A loan authorisation system, ECLAS,
demonstratesthe feasibility of the cognitive framework,the
significanceboth of the efficiencyof systemperformance
and the
reductionin the calculationof the case matches.
Keywords.case-based-reasoning,rule-based-reasoning, expertsystems,case-match-reduction.
Introduction
Case-based reasoning (CBR) uses periodic experiences
stored in cases as a basis for decisions and has been
implemented in a wide range of fields (Kolodner 1993;
Aamodt & Plaza 1994; Watson 1997). A large number of
cases are used in order to increase confidence in the
solution. The more cases gathered, the better that past
experience is utilized. Manycorporations develop largescale or multiple case-based systems. In such situations
the case base load may becomehuge. Inevitably this has
stimulated research into increasing problem solving
efficiency in large case-based problems. Researchers who
focus on representation aim to organize and classify the
case base in order to efficiently point to an appropriate case
(Kolodner 1983a; 1983b; Koton 1989; Schank 1982)n. For
example, Category-ExemplarModel(Bareiss 1988, Porter,
Bareiss, & Holte 1990)~ structured the case base in a
semantic networkby using categories, indexes pointers and
cases. Similar cases are assembled under a category.
t Theoriesof Bareiss1988;Kolodner
1983a;1983b;Koton1989;Porter.
Baxeisss.&Holte1990;Schank
1982drawnfromWatson
&Marir1994;
Aamodt
&Plaza 1994.
33
An index maypoint to a case or a category. The indexes
maybe features linked from problem descriptors to a case
or a category, from categories to their cases or from
categories to the adjacent cases. The well inter-linked
networkefficiently supports case matchin case retrieving.
Others, workingon similarity assessment attempt to
efficiently cluster fewer cases as the initial partial match
(Kolodner 1993). A measure of the difference between
two cases is calculated; cases with the smallest difference
are considered most similar. The difference is the sum of
the distance of features between the new case and one
retrieved case. A numberof strategies strive to provide
increasingly close estimation of differences betweencases.
Yet other approachesfocus on constraint satisfaction
problems aiming at reducing the amount of conflict
checking and, thereby reducing the numberof new choices
provided. Huang & Miles (1996) feature a case-based
travel reservation system embeddinga rule-based modelto
specify constraints before reasoning with the case base.
For example,a special offer is only available for a specific
duration in a non-peak season. The special offer is a
static constraint which specifies duration and price of a
tour. The dynamic constraints are features changed over
time such as customer requirements, number of people,
tour duration, flight, and accommodation...etc.. These
dynamicconstraints, e.g. accommodation,are specified by
the static constraint, e.g. hotel in grade 4, before recalling
cases from the large case base. The heuristic limits the
number of recalled cases. In the above-mentioned
approaches, system efficiency is ameliorated from one to
another, but the problem of case memorydoes not come
under consideration.
Proposed System
The case match reduction approach suggested here
proposes to increase the speed of case-based reasoning by
effectively reducing case memory. In the traditional
¯ scenario, CBRcarries out an initial mappingof the new
case onto the prior cases to generate a cluster of similar
cases and then; keeps comparingthe similarities of these
cases to those of the newcase. This results in a geometric
increase of the comparisonsof the similarities of features.
The case matchreduction alternative is therefore suggested
in order to reduce the number of comparisons to each
similar case in the mapping process. The suggested
process reduces the cluster of the comparisonsrequired for
new cases. It greatly reduces the complexity of the case
matchand increases the efficiency of the matchingprocess.
Since CBRis a cognitive oriented technique, the
suggested procedure is designed in terms of cognitive
psychology. Objective knowledge is distinguished from
subjective knowledge. Objective knowledge is logical,
explicit, and rational (has a definite answer). In contrast,
subjective knowledgeis implicit, uncertain, and imprecise
(has a complex inter-relationship).
The approach
represents objective knowledge in a logical rule-based
reasoning process ’and subjective knowledgein case-based
reasoning one.
Definition of the Objective Characteristics.
1. Strong theory domain: These statements (such as
financial principles or banking regulations) can be proved
true or false (Aamodt1994). For example, the total net
worth of a corporate applicant is lower than the bank’s
aggregate lower limit in a loan granting system. The size
of the company is too small to be accepted. Thus the
loan application obviously will not be approved.
2. Rationality:
Procedures or algorithms that
provide unequivocal results (a computational function or
mathematical model). Applying these rules or procedures
producesrational as opposedto subjective or bias decisions
(Brown 1987). For example, Since 1968, the Z-Score
model has been developed as an index of financial
vulnerability by using a combinationof traditional financial
ratios and statistical techniques. The equation involves
five standard categories of corporate performance, which
are given relative importanceby different weights, to find
the discriminatingvalue of failure.
3. Explicitness : This criteria is complementary
to the
strong theory domain. The behavior of some properties
cannot be adequately estimated, but the influence can be
explicitly known by some specific contexts or can be
definitively distinguished. For example, the results of the
influence can be represented by tendency, e.g. good,
indifferent, and bad. For example, if the Z Score is 1.8 or
less, the result wouldsignify a very higher probability of
insolvency. If the Z score is between 1.81 and 2.99, the
probability of failure cannot be certain. Whenthe Z score
is 3.0 or higher, the companyis unlikely to fail (Altman
1968). Therefore, whenthe Z score is 1.8 or less, the result
can be determined as unacceptable. A primary decision will
be derived. Whenthe Z score is higher than 1.8, the
results will be further investigated.
4. Direct influence on the final solution : The
objective features should be features directly influencing
the final solution. If features are explicit, rational, or have
strong theory but do not havea direct influence on the final
decision, the features are not categorized as objective
knowledge.For example, current ratio has a strong theory,
a value of 1 normally indicates good financial health.
However, current ratio is combined with quick ratio,
receivables, payables, and turnovers to define a criterion of
liquidity, whichforms the feature whichhas a direct effect
on the final decision. Thus current ratio is not an
objective feature in this formulation.
Definition of the Subjective Characteristics.
1. Weaktheory domains: Statements are moreor less
critical, and more strongly or moreweaklybelieved, rather
than proved true or false (Aamodt 1993). For example,
financial leverage can only be identified as "the less, the
better". There is no determining value for good or bad
leverage.
2. Uncertainty, unknownor anomaly : Knowledgein
the criteria usually can be inferred and explained by the
surrounding contexts or the features under the same
criterion. For example, when a feature of liquidity is
unknown,the features in the criterion of liquidity such as
inventory, receivables, payables can be usedto replace the
mainfeature of liquidity.
3. Semantic relationship : The representation
paradigm in this criterion is not primarily a syntactic
problem, but one of a semantic structure (Aamodt1993).
Understanding this inter-relationship
is important in
obtaining a final solution. For example, the quality of
liquidity refers to the quality of receivables, of payables, of
turnover, and other factors. The quality of receivables
depends on days of receivables and percentage of sales to
receivables. Likewise, the quality of payables dependson
days of payables and percentage of sales to payables.
Similarly, the quality of turnover depends on days of
turnover and percentage of sales to turnover.
Input data
Objective knowledge
In
Rule-Based Reasoning
Subjective knowledge
In
Case-Based Reasoning
Proposed
Solution
j
Figure1 SystemIntegration
Integration. The proposed method merges the rule-based
system and the case-specific knowledge system into an
integrated system (Figure 1). The rule-based process
initially infers objective knowledge and determines a
primary decision. If objective knowledgeis sufficient to
draw out a final decision, the primary decision will be
34
determined. For example, a credit card application will be
denied right awayif a prior record of a substantial unpaid
balance within the past six months is found for the
applicant. In some cases, it will not be necessary to
perform further analysis. If objective knowledgepasses
the primary examination in the rule-based reasoning part of
the system, a more in-depth investigation
will be
undertaken in the case-based reasoning part. The casebased reasoning procedure continues to infer subjective
knowledgeand generates a final conclusion. For example,
if the credit card applicant meets the minimumincome
requirements, e.g. the minimumrequirement for a gold
card is a yearly income of GBP15,000(USD25,000), and
has no prior record of unpaid balances or other such history,
the application would be provisionally accepted. The
credit union still needs to investigate the applicant’s
monthly financial burden, e.g. car loan, housing loan, or
other credit card payments. The features of the monthly
financial burden cot~Id be examinedusing the case-based
reasoning part of system. The case-based reasoning
process infers the features of subjective knowledgeand
concludesif the applicant is qualified to receive a credit
card. The rule-based process and the case-based process
therefore are vertically integrated.
Loan Authorization
System
A loan authorization system is used to demonstrate the
feasibility of distinguishing objective and subjective
judgements. The domain knowledge is acquired from a
globally centralized credit analysis process in a French
bank. The branch evaluated only lends to corporate
customers. A credit control procedure summarizes the
figures of last year in the credit proposal into eight
financial criteria which depict the company’s debt
serviceability. Each criterion takes a value of 0 or 1.
The total score marksfor the criteria range from 0 (worst)
to 8 (best) on the financial health of a company.
1. Total Exposure:
Lender’s total exposure/tangible net worth < 50%= 1
2. Gearing1 :
(STL2+bonds)/tangiblenet worth < 3 (times) = 1
3. Gearing 2 :
3 + off-balance sheet
(Gearing 1 + LTD
activities4)/tangible net worth < 4 (times)
1
4. FEYFN"
Financial expenses/turnover < 6% = 1
5. LTD/CF:
1
Long term debt/cash flow < 5 (years)
6. WorkingCapital :
Working Capital > 0 (positive) =
7. STD/MTN:
ST borrowings/monthly turnover < 5 (times) =
8. STLAJARI:
(ST borrowings - Liquid Assets) / (Account
Receivables + Inventory) < 70% =
Determining Objective and Subjective Features
The demonstration system determines the objective and
subjective features based on these eight criteria. The sum
of the eight criteria evaluates total performance of the
company which has a direct impact on the loan
authorisation. For example,if the score is only 2, the loan
application normallywill not be accepted.
Total Exposure : The maximumlending amount
should be no more than 50%of the borrower’s net worth.
This assesses the value of the company’sequity from the
point of view of risk. In the case of bankruptcy, will the
amountof equity be sufficient to repay the loan to the
lender ?
Gearing 1 and Gearing 2 : These evaluate the
borrower’s net worth in order to assess whether they will
be able to repay the borrowed principal. The larger the
borrowed sum, the greater the risk for the lender.
Gearing1 discusses the ability to pay the short-term debts,
including short-term loans and bonds, in net worth terms.
Extending Gearing 1, long-term borrowings, occasional
expenditures in the off-balance sheet are also considered in
Gearing 2. The short-term borrowings are acceptable
whenthey are less than 3 times the borrower’s net worth,
and all borrowings together with the short-term and longterm debts must be less than 4 times. For instance, for
Gearing1 if the ratio is 1.5 then Gearing1 obtains a score
of 1. The repayment of short-term loans by the company
using net worth can be guaranteed.
FE/TNand LTD/CF: FE/TNestimates the ability to
repay the interest expenses using the company’sturnover
in the short term. For a wholesaler bank, the interest is
always a large amount at each drawndown,e.g. monthly
interest could be as substantial as GBP2,000for a working
capital borrowing in the amount of GBP2,000,000from a
securities house. Whenfinancial expenses and turnover
both increase and the score of FE/TNstill is 0, the situation
should be further discussed. For example, increasing
turnover also makes the interest expenses increase. In
fact, it is goodto increase turnover. If turnover increases,
S decrease, the profitability and
but net result and EBITDA
the solvencyare still suspicious.
LTD/CFassesses the borrower’s ability to annually
compensate the debts. The greater the cash flow, the
stronger the company’s ability to meet payments.
LTD/CF
illustrates howmanyyears cash flow will take for
the long-term debts. The ratio of LTD/CF
uses the idea of
the amortization of depreciation to consider amortizing the
long-term debts using cash flow. For example, the ratio
of LTD/CFis 2. It means the cash flow of the applicant
5 EBITis an abbreviate of Earning Before Interest and Tax.
2 ST=Short-term, STL= Short-term loans.
3 LTD= Long-term debts.
4 Activities not included on balance sheet e.g. unreported expenditures.
EBIT= Net result + Financial interests + Taxes.
EBITDA
= EBIT + Depreciation Amortization.
35
can repay the long-term debts in 2 years.
Working Capital, STD/MTNand STLAJARI: The
three features explore the liquidity in the short term. When
WorkingCapital is negative, the score for this component
is zero. However, as long as the STD/MTNand
STLAJARIratios are satisfactory,
then liquidity is
considered acceptable. It means monthly turnover can
still service short-term debts and even if liquid assets are
not considered, the short-term debts can be recovered
through the company’saccount receivables and inventory¯
If the trend of WorkingCapital is declining, other factors in
the criteria of STD/MTN
and STLA/ARI,like inventory,
receivable and payables should be considered before
makinga decision.
When Working Capital is positive: if STD/MTN
obtains a score of 0, it indicates that the short-term loans
are more than 5 times the profitability
of turnover.
Therefore, the decision maker should look into the
¯ relationship betweenthe loans maturedwithin one year and
turnover which is presented in the item of Treasury
Pass.6/Turnover. The item of Treasury Pass./Turnover
shows the numberof days in which turnover can be used to
repay the loans matured in one year. The greater the
numberof days, the moredifficult turnover can recover the
loans maturedin one year.
WhenWorking Capital is positive but STLA/ARI
gets a score of 0, it showsthat the short-term loans minus
liquid assets are higher than 70%of the account receivables
and inventory. Ratios like inventory/turnover, (inventory
+ receivable)/turnover, and suppliers/turnover provide the
decision makerwith sufficient information to have a better
appreciation of the liquidity situation. The smaller the
three ratios represent, the better the workingcapital is put
to use~ For instance, a manufacturer’s turnover can pay
inventory in 80 days, pay inventory plus receivables in 90
days and suppliers in 75 days. The liquidity can last
around 80 days and wouldbe acceptable in a short term.
Total Exposure, Gearing 1, and Gearing 2 can be
seen as banking regulations from head office. Headoffice
has precise standards on the first three criteria for a loan
applicant. They can be judged definitely since they can
be determined precisely from the answers and they have a
direct influence on the final decision. WorkingCapital
can also be identified as influential in the final decision.
Despite WorkingCapital having an effect on the seventh
and eighth criteria, it has a strong influence on primary
decision-making. The Headof Credit indicates that Total
Exposure, Gearing 1 and Working Capital are the most
important features amongthe eight criteria. They will not
look further if a loan applicant concurrently fails to pass
the requirement of the first three features. The four
criteria, Total Exposure, Gearing I, Gearing 2 and Working
Capital are identified as the objective features.
On the other hand, FF_dTN, LTD/CF,STD/MTN
and
STLA/ARI
are determined as the subjective features. The
four criteria have a complex relationship with other
6 Treasury Pass = Short-term loans + Current portion of long-term debts.
financial factors¯ For example,FE, fI’N is related to net
result and EBITDA.LTD/CFis associated with the net
result,
STD/MTNto Treasury Pass.fl’urnover
and
STLA/ARIto the inventory, receivables and payables.
The four criteria mayall be subject to exceptions from time
to time. The humanexpert must use subjective judgement
to examinethe situations whilst consulting the four criteria.
Expert Case-based Loan Authorisation
(ECLAS)
System
~
I Initial
ns
I
(SG Limit ! net worth < 50%)
[TotalExposure
]
(Working capital
> 0)
[W°rkinl~Capital
(Gearing 1 < 3)
>=0
~
i
=3
I
i
(Gearing 2 < 4)
Figure
in the
<0
[
..°
[Gearing2
<50~ I
I
J
[
<4 I
>=4
I
[
2. Part
Rule-Based
of
Remarks:
u =underlimit
up =positive
underlimit
St =Short-term
Lt =Long-term
s
=safe
ss =a safe
short-term Ioar
Knowledge
Base
Reasoning
Process
The Headof Credit ranks the significance of the features
on lending decisions as Total Exposure, WorkingCapital,
Gearing 1, Gearing 2, PT_/I’N, LTD/CF,STD/MTN,and
STLAJARI.The former four features are dealt with using
objective judgements and the latter four by subjective
judgements. The loan authorisation
system, which
integrates rule-based reasoning and case-based reasoning
procedures, is modeledin a powerful expert system shell,
7.
called wxClips
The knowledgebase in the rule-based reasoning is
partitioned
into four layers - TotalExposure,
WorkingCapital, Gearingl and Gearing2 in the sequence of
their importance(Figure 1). A hierarchical tree determines
an appropriate branch whenthe system receives an answer
from the user. For example, the first question asks "How
muchis the total exposure requested ?". If the answer is
less than "50" (percentage), the system chooses the left
branch and determines"underlimit" as a current state of the
layer of "TotalExposure" and the system’ prunes another
sub-tree. Next, the second question "Howmuch is the
working capital of the company?" is posted out. Given
that the amountof workingcapital is greater than "0", the
system points to the left branch. Once "positive-u" is
selected as the state of the layer of "WorkingCapital".
Thus, the system asks "How much is Gearing 1 of the
company ?". Suppose the value of Gearing 1 is not
7 Homepage of WxClips:http://web.ukonline.co.uk/julian.smart/wxclips.
36
greater than "3" which satisfied the system requirement,
"St-s-up" is chosenas the state of the layer of "Gearing1".
The last question is "How much is Gearing 2 of the
company?". If the value of Gearing 2 is less than "4", the
system moves to "Lt-s-ss-up". Once the system finds out
enough information to extract a conclusion, a primary
decision is achieved. In this case, the loan application is
initially approved and it is allowed to go for further
examination in a CBRprocess. On the other hand, if the
system reach a primary decision of "rejected",
the
investigation of the loan application will cease at this stage.
The ECLASsystem integrates the rule-based part
system and a case-based reasoning part. Rule-based
reasoning filters out the failure features as a primary
decision (Figure 2). Only applicants who satisfy the
requirements of the primary decision will be considered for
further analysis using subjective decision-making.The loan
applicants failing the investigation of the primary decision
are dropped now.
unsatisfactory
~
satisfactory
objective judgement
unsatisfactory
satisfactory
Figure 3. Integration
System
The cases to be examinedin the case-based reasoning
system are divided into two categories: Authorised - the
bank agrees to the loan application with reference to a good
companyperformance. Denied - applicants in this category
are considered highly risky due to their poor financial
positions.
The loan authorisation uses a case-based reasoning
technique based on the nearest-neighbor matching method
(Kolodner 1993; Watson 1997). The distance score
represents howclose a base case is to the target case. The
smaller distances signify better matches and the one with
the minimum
distance is recognized as the nearest neighbor.
The case base is serially indexed and each case is
composed of the four subjective features. The expert
allocates the features weights from 4 (mostsignificant) to
(least) with respected to the influence of the investigated
conclusion.
The sequence arranged from the most
weighty to the least is b-F_ZI’N, LTD/CF,STD/MTN
and
STLA/ARI.The system first and foremost tallies distance
for each of the features on a scale from 0 to 5, with 0
symbolizing no difference and 5 symbolizing maximum
difference. After working-out the individual distances,
the system weights those distances by the relevant
weightings and averages them out to find an overall
distance. The formula (Sinha, & Richardson 1996) the
systemuses is:
37
n
d
=
~ Wi* dista" (target, case base)
i=1
where
d
= distance of case base from target
Wi
= weight of feature i
disti (target, casebase)
= distance of case base from target on feature i.
If only one minimumcase is retrieved, the system
next makes an authorizing decision based on the outcome
of the neighbor. If the nearest neighbors are a set of cases,
the decision follows the outcome in accordance with a
majority of results from morethan 50%of the neighbors.
The similarities assessment needs at least 1 rule to
count the distance from the new case to the case base using
a most elementary distance calculation method, e.g. the
nearest-neighbor matching method shown in the ECLAS
system. The distance of each feature in a case is
estimated using the distance formula once. ECLAS
separates 4 objective features from the case-based
reasoning system. Therefore, 20,000 calculations are
avoided for features in a case base containing 5,000
examples. This simultaneously reduces by one half the
memorystorage requirement from 40,000 features of the
case base to 20,000. The system performance verifies the
efficiency of the approach. The proposed case match
reduction technique appears to geometrically reduces the
numberof the features to be calculated whenthe numberof
cases increases.
Conclusion
Case-based reasoning stores past experiences on cases to
draw resolutions. In practice the number of eases used
should be quite large in order to win credence with users.
In this situation case memorywill become massive. The
case match reduction method proposed is based on a
cognitive hypothesis. The hypothesis suggests that the
objective and subjective knowledgeshould be represented
by different techniques.
For example, logicality,
simplicity and ability of obtaining an exact answerin rulebased reasoning are appropriate for objective knowledge
which is dealt with in the RBRcycle. Onthe other hand,
case completeness, dynamismover time and the ability of
case justification
are more suitably represented by
subjective knowledge in the CBRcycle. Implementing
the representation schemeof rule-based reasoning requires
comparatively speaking less memorythan that of casebased reasoning, which in turn will allow for reduction of
total memorystorage.
The feasibility of implementingthe objective features
in the rule-based reasoning and the subjective features in
the case-based reasoning has been demonstrated in the
ECLAS system. The number of calculations
are
estimated to be reduced by 20,000 for 5,000 case
comparisons whenthe four features employedin the casebased reasoning are movedto the rule-based reasoning.
When the number of case comparisons and features
increases, the number of calculations
and storage
requirement will geometrically decrease and the benefits of
the proposed approach increase.
Watson, I., & Marir, F. 1994. Case-BasedReasoning: A
Review. The KnowledgeEngineering Review 9(4). Also
available on internet at
http://www.ai-cbr.org/classroorn/cbr-review.htm#CaseBased Reasoning Techniques.
Watson, I. 1997 Applying Case-Based Reasoning. Morgan
KaufmannPublishers, Inc.
References
Aamodt, A. 1993. Explanation-Driven Case-Based
Reasoning. In Topics in Case-BasedReasoning,edited by
S. Wess, K. Althoff, M. Richter. Proceedings of First
European Workshop, EWCBR93,
Kaiserslautern, Germany.
pp. 274-288.
Aamodt, A. 1994. A Knowledge Representation System
for Integration of General and Case-Specific Knowledge.
Proceedingsof lEEEConferenceon Tools for Artificial
Intelligence.
Aamodt,A., & Plaza, E. 1994. Case-Based Reasoning:
Foundational Issues, MethodologicalVariations, and
System Approaches. AI Communications7(i): 39-59.
Alsoavailable on internet at http://
www.iiia.csic.es/People/enric/AICom.html#RTFToC1.
Appendix
1. Integration name/category:
Rule-based reasoning (RBR)and case-based reasoning
(CBR)procedures.
2. Performance Task:
Retrieval. A demonstration is built in the banking
loan authorization domain.
Altman, E. 1968. Details drawn from Failure Prediction
Model(1996) Presentation Notes of Cooperate Decline
and Turnaround,in the evaluated bank.
Bareiss, E.R. 1988. PROTOS:
A Unified Approach to
ConceptRepresentation, Classification, and Learning,
Ph.D. thesis, Dept. of ComputerScience, University of
Texas, Technical Report CS 88-10, Dept. of Computer
Science, Nashville, TN:Vanderbilt University.
Brown, H. I. 1987. Objectivity. Observation and
Objectivity. OxfordUniversity Press Inc., 191-230.
Huang, Y., & Miles, R. 1996. Using Case-Based
Techniquesto EnhanceConstraint Satisfaction Problem
Solving. Applied Artificial Intelligence: 307-328.
Kolodner, J.L. 1983a. Maintaining Organization in a
Dynamic Long-Term Memory. Cognitive Science (7).
Kolodner, J.L. 1983b. Reconstructive Memory,a
Computer Model. Cognitive Science (7).
Kolodner, J. L. 1993. Case-based Reasoning. Morgan
KaufmannPublishers, Inc.
Koton, P. 1989. Using Experience in Learning and Problem
Solving, MassachusettsInstitute of Technology,
Laboratory of ComputerScience (Ph.D. thesis, October
1988). M1T/LCS/TR-441.
Porter, B.; Bareiss, R.; &Holte, R. 1990. ConceptLearning
and Heuristic Classification in WeakTheory Domains.
Artificial Intelligence, 45(1-2): 229-263.
Schank, R. 1982. DynamicMemory; A Theory of
Reminding and Learning in Computersand People.
CambrigeUniversity Press.
Sinha, A., & Richardson, M. A. 1996. A Case-Based
Reasoning System for Indirect Bank Lending. Intelligent
Systems in Accounting, Finance and Management(5):229240.
38
3. Integration Objective:
A case match reduction approach. Representing
features in RBRrequires comparatively speaking less
memorythan that in CBR,which in turn allows for a
reduction of total memorystorage. The integration of
RBR and CBR improves system efficiency,
case
memory economy and user comprehension of
explanations.
4. Reasoning Components:
RBRfor pre-screening, followed by CBR.
5. Control Architecture:
Sequential.
6. CBRCycle Step(s) Supported:
Pre-processing in RBRand then retrieval in CBR.
7. Representations:
Representing objective features in RBRand subjective
features in CBR.
The proposed method suggests that objective and
subjective knowledge should be represented by
different techniques. For example, RBRis appropriate
for objective knowledge (logicality, simplicity and
ability of obtaining an exact answer.) CBRis used to
represent subjective knowledge.
8. Additional Reasoning Components:
Nil.
9. Integration Status:
Demonstrator
of an Expert Case-based Loan
Authorization System (ECLAS), which demonstrates
the feasibility of distinguishing objective and subjective
judgements.
10. Priority future work:
Empirical evaluation. The demonstration system takes
into account only eight typical features. It is intended
to extend the core model ECLASto a full-scale
production version by means of other complementary
sub-dimensions.
Download