■ This article presents an overview and survey of cur-rent work in case-based reasoning (CBR) integra-tions. There has been a recent upsurge in the inte-gration of CBR with other reasoning modalitiesand computing paradigms, especially rule-basedreasoning (RBR) and constraint-satisfaction prob-lem (CSP) solving. CBR integrations with model-based reasoning (MBR), genetic algorithms, andinformation retrieval are also discussed. This arti-cle characterizes the types of multimodal reason-ing integrations where CBR can play a role, identi-fies the types of roles that CBR components canfulfill, and provides examples of integrated CBRsystems. Past progress, current trends, and issuesfor future research are discussed.

Case-based reasoning (CBR) is an AI para-digm that can be synergistically com-bined with other approaches to facili-

tate a broad array of tasks (Aha and Daniels1998). This article presents a brief introductionto CBR, a review of other approaches withwhich CBR has been combined, an overview oftasks CBR integrations can perform, a discus-sion of open issues in CBR integration, and alook at synergies achieved through CBR inte-gration. It characterizes the types of multi-modal reasoning integrations where CBR canplay a role, identifies the types of roles thatCBR components can fulfill, and providesexamples of integrated CBR systems.

CBR is the process of using past cases tointerpret or solve new problem cases (Kolodner1993; Riesbeck and Schank 1989). Excellentextended introductions to CBR are available(Aamodt and Plaza 1994; Kolodner and Leake1996); therefore, this overview is brief. Centralto CBR is the determination of which past

cases are similar enough to the new case towarrant consideration for reuse in the newproblem. There are two basic kinds of CBR: (1)interpretive and (2) problem solving. Interpre-tive CBR involves use of past cases, typicallycalled precedents, to create an analysis and jus-tification for the interpretation of a new case.Problem-solving CBR involves adapting a solu-tion to a past problem, typically a design orplan, to meet the requirements of the new sit-uation. A lawyer arguing a case, for example,would use interpretive CBR, but an engineerconfiguring a motor would use problem-solv-ing CBR. Problem-solving CBR can be furthersubdivided into transformational and deriva-tional CBR. In transformational CBR, past solu-tions are used directly. In derivational CBR, theproblem-solving processes by which solutionsare derived are reused.

All CBR follows the same basic steps: (1) ana-lyze the new case (or problem); (2) based onthis analysis, retrieve relevant past cases froma case base; (3) based on a “similarity metric,”rank retrieved cases according to how relevantor useful they are with respect to the new case,and select one or more “best” cases to use insolving the new case; (4) create a solution tothe new case; (5) test and explore the proposedsolution; and (6) if appropriate, add the newcase and its solution to the case base and indexit so that it can be retrieved for future use.

Critical ingredients in CBR are means toindex and retrieve cases from the case base,methods to assess similarity, methods to com-pare and contrast cases, and methods to createsolutions. Solution methods in the two basictypes of CBR are quite different: In interpretiveCBR, the methods produce interpretationswith supporting rationales, and in problem-solving CBR, new plans or designs are pro-

Articles

SPRING 2002 69

Case-Based ReasoningIntegrations

Cynthia Marling, Mohammed Sqalli, Edwina Rissland, Hector Muñoz-Avila, and David Aha

Copyright © 2002, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2002 / $2.00

AI Magazine Volume 23 Number 1 (2002) (© AAAI)

open issues in CBR integration follows. Thearticle closes with a summary and conclusions.

Case-Based–Reasoning Integration Techniques

In this section, approaches and techniques forintegrating CBR with other reasoning modali-ties and computing paradigms are presented.

Case-Based Reasoning and Rule-Based ReasoningRBR was the first modality to be successfullyintegrated with CBR. RBR is the reasoningapproach behind classical expert systems,including MYCIN, which diagnosed infectiousblood diseases; XCON, which configured VAXcomputers for the Digital Equipment Corpora-tion; and PROSPECTOR, which determined whereto search for mineral deposits (Hayes-Roth,Waterman, and Lenat 1983). An RBR systemcontains a rule base, or a set of if-then rules; aninference engine, or a means of applying therules to solve problems; a working memory, ora means of maintaining the current problemstate; and frequently, an explanation facility,or a means of showing a user the sequence ofrules that led to a conclusion.

Each rule represents a small piece of knowl-edge that can be combined, or chained togeth-er, with other rules to infer conclusions orderive solutions to problems. RBR differs fun-damentally from CBR in both knowledge rep-resentation and knowledge utilization. Someimportant distinctions noted in Kolodner(1993) are that cases represent larger chunks ofknowledge than rules, and they compose thisknowledge in advance of run-time inferencing;rules represent patterns, but cases are con-stants; and rules are applied when they matcha situation exactly, whereas cases can be usedin partially matching, or similar, situations.

Some rationales for integrating CBR withRBR are that problem domains might naturallycontain both rules and cases; human reasonersmight naturally use both rules and cases; therules of a domain might be vague, or open tex-tured, and therefore require interpretation;rules can have exceptions that must be han-dled; rules can enable evaluation of solutionswithout supporting initial generation of solu-tions; and cases can be few or difficult toacquire in a domain.

The earliest CBR-RBR hybrids were built instatutory legal domains. These domains natu-rally include rules, or statutes, and cases, orlegal precedents. A famous hypothetical rulefrom legal philosophy illustrates the problemsof statutory interpretation (Hart 1958). Con-

duced. Interpretive CBR relies heavily onmethods to reason analogically with similari-ties and differences between cases. Good exam-ples of interpretive CBR can be found in thelaw. The HYPO system was one of the earliestand best examples of an interpretive CBR sys-tem (Ashley 1990; Rissland, Valcarce, and Ash-ley 1984). HYPO modeled how lawyers arguecases involving trade secrets law by presentingpoints and counterpoints based on cases previ-ously settled in courts of law. Problem-solvingCBR relies on methods to adapt cases andassess how modifications impact the overallsolution (for example, does satisfying a newgoal undo satisfaction of an old one). One ofthe earliest and best examples of problem-solv-ing CBR was CHEF (Hammond 1989, 1986). CHEF

was a planner, which created new plansthrough adaptation of old ones. A plan in CHEF

was a recipe for preparing a complex dish.Note that a system that only analyzes and

retrieves cases is not a complete CBR system.However, in many hybrid systems, it is exactlythe early steps of CBR—analysis, retrieval,ranking, selection—that are often used in con-junction with another reasoning method. Oth-er, more complex, hybrid systems use full-scaleCBR. One might distinguish between these sys-tems with the use of a small or a big “r”; thatis, CBr or CBR. In this article, the primary focusis on CBr (small r) hybrids.

Although CBR is a powerful paradigm in itsown right, there are strong motivations forintegrating it with other reasoning modalitiesand computing techniques, including rule-based reasoning (RBR), constraint-satisfactionproblem (CSP) solving, model-based reasoning(MBR), genetic algorithms, and informationretrieval. Researchers have reported that inte-grated approaches have enabled them to moreaccurately model the knowledge available in aproblem domain, compensate for the lack of apredictive or complete model of a problemdomain, simplify the knowledge-acquisitionprocess, improve solution quality, improverun-time efficiency, leverage past problem-solving experiences, and compensate for theshortcomings of one approach by capitalizingon the strengths of another (Aha and Daniels1998; Rissland and Skalak 1991).

In summary, there are many reasons for inte-grating CBR with other reasoning modalitiesand many ways to do so. In the next section,different approaches to CBR integration arepresented. Then, tasks that have benefitedfrom CBR integration are described, includinginterpretation and argumentation, design andsynthesis, planning, and the management oflong-term medical conditions. A discussion of

Articles

70 AI MAGAZINE

sider the rule “No vehicles in the park.” Part ofthe problem in administering this rule is deter-mining what counts as a vehicle. Clearly, a caror a bus is a vehicle, but is a motorized bike avehicle? A motorized wheelchair? A nonmotor-ized wheelchair? Further, how ironclad is thisrule: Does it really mean never, or are theresome situations when a vehicle is allowed? Forexample, how should the rule be applied to firetrucks, ambulances, or park maintenancetrucks? Does the rule really mean “No vehicles,except for emergency vehicles or park mainte-nance vehicles, are allowed in the park”?Should the rule always permit such vehiclesaccess? What about a park maintenance truckthat only drives in the park so its crew can havea picnic lunch? No matter how carefully craft-ed such rules are, there will always be interpre-tive questions about their meaning and scopeand their constituent terms. The law deals withthese problems by using cases in conjunctionwith rules. Cases can stand for implicit updatesto the rules, for example, by registering unspo-ken exceptions (emergency vehicles) andrequirements (on official business). Cases allowa rule-based regime to adapt to its changingenvironment without necessarily requiring therules to be rewritten at every new turn ofevents.

Good examples of hybrid CBR-RBR legal sys-tems are CABARET (Rissland and Skalak 1991),GREBE (Branting 1991), and IKBALS (Zeleznikow,Vossos, and Hunter 1994). The CABARET systemoperates in the area of U.S. tax law concerningthe home-office deduction. It integrates CBRand RBR—tax cases and tax regulations—byusing a rule-based agenda mechanism. Basedon the status of tasks in progress by CBR andRBR modules, heuristic rules are used to postand prioritize tasks on an agenda of tasks. Thisseminal system is described in more detail lat-er. GREBE operates in the area of Texas employ-ment law concerning worker disability forinjuries sustained during the course of employ-ment. It uses CBR and RBR to determine andjustify the legal conclusions of its cases. UnlikeCABARET, it does not use an agenda to mediatebetween strategies. Rather, it always tries bothstrategies to find the best justification for aconclusion. IKBALS works in two Australian legaldomains: (1) worker disability and (2) lendingby financial institutions. RBR is its dominantreasoning mode. CBr is used to retrieve appro-priate cases for review by lawyers when rulesproved inadequate to derive a solution. IKBALS

also integrates information retrieval, givingusers access to sources such as legal treatises inaddition to statutes and cases.

Another early system that capitalized on

CBR-RBR synergy is ANAPRON (Golding andRosenbloom 1991). ANAPRON is a speech synthe-sizer that pronounces American surnamesaloud. Because American surnames originate inmany different languages and become Ameri-canized in many different ways, ANAPRON’s taskis complex. ANAPRON uses a CBR module toimprove the accuracy of an essentially RBR sys-tem. It uses rules to generate a probable pro-nunciation for a name and then uses cases tohandle exceptions to the rules. Although thisexample is not from Golding and Rosenbloom(1991), one might think of the exception madefor the popular British television characterHyacinth Bucket, who insists that her surnamebe pronounced bouquet. Numerous exceptionsarise as people try to stand out or fit in or deferto whatever their neighbors happen to callthem. Such cases might be viewed as extremelyspecific rules. For example, if you happen tomeet Hyacinth Bucket, then remember toaddress her as Mrs. Bouquet. Golding andRosenbloom reported that rules and cases hadcomplementary strengths in ANAPRON. Ruleswere good for capturing large trends, and caseswere good for compensating in situations notcovered by the trends. Without CBR, ANAPRON’saccuracy might have been improved throughcontinued refinement of the rule set. However,with over 600 rules in the system, a point ofdiminishing returns had been reached. It waseasier to integrate cases than to formalize morerules.

CBR and RBR have also been combined toperform the task of harmonizing melodies(Sabater, Arcos, and López de Mántaras 1998).This task has traditionally been approached ina rule-based manner because there are well-established harmonization rules that reflectcommon musical practice. However, theserules are not constructive in nature. As Sabateret al. explain, “... the rules don’t make themusic; it is the music which makes the rules”(p. 147). It is easier for people to appreciategood harmonization or detect bad harmoniza-tion than it is for them to generate good har-monization in the first place. GYMEL integratescases, which are musical phrases from Catalanfolk songs, with general harmonization rules.It inputs a musical phrase and outputs thesame musical phrase with a set of accompany-ing chord sequences. CBR is the dominantmode of reasoning, and RBR is applied whenCBR can not provide a solution. Althoughexpanding GYMEL’s case base might allow CBRto provide more solutions, Sabater et al. didnot find it easy to formalize a large number ofmusical phrases as cases. They believe thattheir approach is applicable to other domains

Articles

SPRING 2002 71

the constraints of the system to be modeled. Inaddition, it can be applied to many differentdomains because of its simple but rich repre-sentation. Constraint-satisfaction problems(CSPs) involve finding values for variables sub-ject to restrictions on which combinations ofvalues are acceptable.

The advantage of CSP is that it is a reasoningmode that provides both modeling and solvingof a problem within the same framework. CSPprovides a simple and convenient way of rep-resenting problems because it is a natural anddeclarative approach to modeling. CSP is alsodomain independent because it can hide manydomain-specific issues and be used at a moreabstract level. When a problem is representedas a CSP, it can be solved independently of theinitial context or domain of application. TheCSP methods are applied to the CSP represen-tation of the problem, which hides the contextused. CSP provides many advanced algorithmsto deal with hard problems (Kumar 1992).Constraint reasoning takes advantage of manymathematical methods and algorithms thatwere improved to deal with CSPs.

CSP uses a representation called a constraintgraph in which the vertexes are variables of theproblem, and the edges are constraintsbetween variables. Each variable has labels,which are the potential values it can beassigned. CSPs are solved using search andinference methods.

The map-coloring problem illustrates howCSP operates. This problem can be stated as fol-lows: “Given a map with N regions borderingeach other and M colors that can be used tocolor each region, determine whether there isan assignment of one color to each region suchthat no two bordering regions have the samecolor.” Figure 1 shows a map-coloring problem.

This problem can be represented as a graph-coloring problem where the nodes representthe different regions, the labels for each nodeare the different colors a region can have, andthe existence of an edge between two nodesindicates that the two corresponding regionscannot have the same color (that is, there is aconstraint of “not equal” between these twonodes). Figure 2 shows the graph-coloringproblem that corresponds to the map-coloringproblem in figure 1.

The graph-coloring problem can be formu-lated as a CSP. The variables of this problemrepresent the regions (X, Y, and Z), the valuesare the different colors (red, blue and green),and the constraint is that no bordering regionshave the same color (that is, no two variablesconnected by a “not equal” edge can beassigned the same value).

where there are an insufficient number of casesand an inadequate rule base for single strategyproblem solving.

Case-Based Reasoning and Constraint-Satisfaction Problem SolvingConstraint satisfaction is a powerful and exten-sively used AI paradigm (Freuder and Mack-worth 1992). It is a natural way of representingproblems because the user needs only to state

Articles

72 AI MAGAZINE

X

ZY

Figure 1. Map-Coloring Problem.

Y

red

bluegreen

red

bluegreen

red

bluegreen

X

Z

Figure 2. Constraint Graph.

The constraint graph of a CSP is the graphicrepresentation of the problem. In this example,the constraint graph of the map-coloring prob-lem is the same as the corresponding graph-coloring problem. The nodes are the variables,the labels are the values, and the edges are theconstraints.

The CSP has a solution if there is an assign-ment of values to variables such that all theconstraints are satisfied. A solution in CSP canmean different things depending on the con-text and the goal to be achieved. The goal canbe to determine whether there is a solution,determine how many solutions there are, findany solution, find an optimal solution, or finda solution with specific characteristics. Figure 3shows one solution of the map-coloring prob-lem in figure 1.

Many other toy problems such as the N-queens problem can also be represented andsolved using CSP, and these problems havehelped in developing methods and tools thatare used in real-world applications. Many real-world applications have used CSP for problemrepresentation and modeling as well as forproblem solving (Puget and Leconte 1995;Wallace 1996). These applications includedesign and configuration, diagnosis, debug-ging, verification, graphics, decision support,scheduling, planning, and resource allocation.

Figure 4 shows how CSP is used to representand solve problems. With the CSP representa-tion of a problem, different methods can beused to solve it independently of the context ofthe application. The two main solving tech-niques are search and inference. There aremany algorithms that use search exclusively,such as backtracking. Backtracking search canrequire exploration of the entire tree of possi-bilities to find a solution. Other algorithmsmake use of an inference method such as arcconsistency. Research and experience haveshown that the most successful techniques forsolving CSPs are the ones that combine searchand inference. The questions, then, are howand when to combine them to get the bestresults. The answers to these questions dependon the domain of application and the availableresources.

The earliest systems to integrate CBR and CSPwere CADSYN (Maher and Zhang 1991) and JULIA

(Hinrichs 1992). CADSYN generates structuraldesigns for buildings. It uses design constraintsfor case adaptation. CADSYN transforms previousbuilding designs to fit new requirements afterconflicts are detected. JULIA designs meal plansfor groups of diners. It uses a constraint propa-gator to identify and resolve constraint viola-tions that arise during meal planning.

The two main approaches to integrating CSPand CBR have been to (1) use CBR to initializethe CSP system and (2) use CSP in the adapta-tion step of CBR. In the first approach, a similarpast case retrieved by CBR is used to positionthe CSP process at a point in its problem spacefrom which to begin its search for a solution;the search travels outwards by means of CSP-based adaptations of the initial case (Sqalli,Purvis, and Freuder 1999). This usage is anexample of CBr because only the retrievalphase of CBR is used. In the second approach,CSP provides CBR with a specific method foraccomplishing adaptation; thus, CSP is aprocess used internally within CBR. The use ofCSP to accomplish adaptation is relatively new.

Weigel’s research prototype (Weigel and Falt-ings 1998) and the COMPOSER (Purvis and Pu1996), CADRE (Faltings 1996; Hua and Faltings1993), IDIOM (Smith, Lottaz, and Faltings 1995)and CHARADE (Avesani, Perini, and Ricci 1993)projects exemplify the two integrationapproaches. The rationale for Weigel’s systemwas that in configuration problems, the usermight have incomplete knowledge of thedomain; so, CBR can be used to provide an ini-tial solution to help the user specify the prob-lem. Even though a configuration domain canallow for the development of a complete con-straint model, customers might not fully beable to describe the products they wantbecause of their incomplete knowledge of theproducts or the complexity of the domain.Weigel’s system proposes an initial solutionthat is close to a customer’s desired productand then modifies it to meet the remainingrequirements. Here, case adaptation is consid-ered to be the process of replacing values in asolution (case) with interchangeable values. Itis easier to adapt a nearly satisfactory, or close-by, solution than to construct a complete solu-tion from scratch. Finally, CSP, in turn, helps increating the case base of initial solutions. UsingCSP to create new cases, as well as using CBR to

Articles

SPRING 2002 73

XY

greenZ

red

blue

Figure 3. Constraint-Satisfaction Problem Solution.

can be architectural or structural and are used toreduce the adaptation space. CADRE formulatescases as instances of particular buildings, andthe constraints restrict the possible modifica-tions that can be made to these cases. There aretwo forms of adaptation: (1) dimensional adapta-tion, where only dimensions of the case arechanged, and (2) topological adaptation, wherethe arrangement and number of spaces andwalls are also modified. Constraints expressingthe admissible modifications are first solved toobtain a constraint system that is as simple aspossible. The complexity of adaptation is thenmanageable so that it can be carried out by theuser in an interactive manner. CADRE has beentested on examples of realistic complexity.

IDIOM is a system for composing apartmentlayout designs using cases (Smith, Lottaz, andFaltings 1995). The goal of the IDIOM project wasto make the solution space easier to explore bydesigners. It supports designers by reducingconstraint complexity and managing designpreferences, thereby restraining proposed solu-tions and further adaptation within feasibledesign spaces. There are three sources of con-straints in IDIOM: (1) the library of cases, inwhich cases are represented as CSPs; (2) theinterpretation of the design by the user; and (3)the domain models. When a case is introducedinto a design, all its associated constraints areadded to the current set of constraints. The usercan then add further constraints to interpretthe case in its new environment. A solution isthen calculated for the layout. Case combina-tion is supported through incrementally solv-ing relevant constraints. Constraints arerestricted to linear and simple nonlinear rela-tionships so that they can be solved rapidly.

help CSP get started, makes Weigel’s integra-tion of CSP and CBR highly synergistic.

COMPOSER solves assembly sequence and con-figuration design problems using a CSP engineas the adaptation mechanism (Purvis and Pu1996). In the COMPOSER system, the motivationwas to achieve domain independence in caseadaptation through CBR-CSP integration.Here, the CBR adaptation process is accom-plished using CSP. Each case is formulated as aCSP, and a CSP repair algorithm is used to per-form adaptation, resulting in a domain-inde-pendent adaptation mechanism. Representingcases as CSPs also allows case combination tooccur naturally. When the case-based reasoneris presented with a new problem, similar casesare retrieved, and a CSP repair algorithm (theminimum conflicts algorithm) is used toachieve adaptation. COMPOSER finds the corre-spondence between an old case and the newone using structure mapping and nearest-neighbor similarity metrics. The minimum-conflicts algorithm is the means by which COM-POSER combines all the solutions to thematched cases into a consistent solution forthe new problem. Because case representationand case adaptation both use CSP, the COMPOSER

system is one of the most thoroughly integrat-ed uses of CBR and CSP. A strong synergyresults because CBR increases the efficiency ofCSP, and CSP helps to formalize the adaptationprocess for CBR (Purvis 1998).

CADRE (Faltings 1996; Hua and Faltings 1993)is used in the domain of floor-plan design. Itworks together with designers and uses case-based spatial reasoning for the design of build-ings. Constraints in CADRE represent numericrelationships among dimension variables. They

Articles

74 AI MAGAZINE

CSP Representation

ProblemStatement Solution

Constraints

Values

Variables

AlgorithmCSP

Figure 4. Constraint-Satisfaction Problem for Problem Representation and Problem Solving.

The CHARADE Project combines CBR and CSPfor situation assessment and planning in forestfire fighting (Avesani, Perini, and Ricci 1993). Itaids decision making in environmental emer-gencies by also incorporating human assess-ment and reasoning. CSP is used to representcases in CHARADE. Constraint propagation tech-niques are applied to part of the plan represen-tation to support adaptation and repair of asector plan. Thus, CSP supports case adapta-tion in CHARADE. CBR supports the goal of pro-viding a quick assessment of an emergency sit-uation, and the constraint-solving processenables the system to determine the bestexploitation of available resources to handlethe emergency.

There have also been other approaches toCBR-CSP integration. ADIOP, a system for diag-nosing interoperability problems in asynchro-nous transfer mode (ATM) networks, uses CBRto compensate for incompleteness in the CSPmodel and to correct and update the CSP mod-el (Sqalli and Freuder 1998). Here, a diagnosticproblem is first modeled as a CSP, and thenCBR is used to compensate for what is missingin this CSP model. The CSP model is incom-plete or incorrect if it is not completely or cor-rectly modeling an interoperability test. ADIOP

can determine incompleteness or incorrectnessby considering similar cases and interactingwith the user. When a model is incomplete orincorrect, ADIOP uses CBR to check if similarcases of model incompleteness or incorrectnessoccurred in the past and to adapt them to thenew situation. Thus, the user is allowed tocomplete or correct the model by updating itusing the new adapted test case. CBR is alsoused to update the CSP model to make it morerobust in solving future problems. Cases pro-vide information for updating the CSP modelthat might change some of its components, forexample, by adding new constraints. An ad-vantage of CSP in this application is that itreduces the number of cases that would berequired by a purely CBR system.

One advantage of CBR-CSP integration isthat CSP-based approaches to adaptation offersophisticated ways to deal with the long-stand-ing problem of case adaptation. Over the years,many researchers have considered adaptationto be the most difficult part of CBR. Some evenfeel that it ought to be avoided altogetherbecause of its complexity (Barletta 1994). View-ing adaptation as a CSP occurred very early inthe evolution of CBR. For example, in the early1980s, Rissland’s CEG (CONSTRAINED EXAMPLE GEN-ERATION) system generated mathematical exam-ples and counterexamples (“ceg’s”) that satis-fied posted desiderata or constraints (Rissland

1980; Rissland and Soloway 1980). CEG did thisby first retrieving past examples that alreadysatisfied many of the constraints and thenattempting to modify these examples to satisfythe rest of the constraints. Although thisretrieval-plus-modification approach workedwell enough when there was little constraintinteraction, the approach was far too simplisticfor many problems. CSP is exactly the sort ofmechanism needed to handle these problems.This system from 20 years ago is not only oneof the forerunners of CBR but a forerunner ofthe mixed-paradigm CBr approach of usingcase retrieval to position further problem solv-ing, such as with CSP.

There are drawbacks and trade-offs, as wellas advantages, in CBR-CSP integration. First,there is overhead involved in accommodatingboth modes in one system, and many addi-tional system components can be required.Second, each reasoning mode has time andspace limitations. CBR might need more spacefor storing cases and more time to performretrieval and matching. CSP might add morecomplexity because of the NP-completeness ofconstraint solving. Third, it might not beappropriate to update CSP models in all prob-lem domains. However, when a balanced inte-gration can be achieved, CBR and CSP comple-ment each other as follows: CSP provides adomain-independent representation of a task;CBR can be used in incomplete domains,where CSP models are difficult or impossible toobtain; CSP provides a knowledge representa-tion that supports problem formulation; CBRprovides a useful learning capability; CSP pro-vides many advanced algorithms to deal withhard problems; and CBR provides CSP with astarting point for using these algorithms. Asynopsis of the basic CBR steps that have beenaccomplished using CSP is given in table 1.

Other TechniquesModel-based reasoning: Model-based reason-ing (MBR) is an approach in which generalknowledge is represented by formalizing themathematical or physical relationships presentin a problem domain. It is perhaps unfortu-nately named in that CSP relies on formalmodels, and RBR can be used to model prob-lem domains in which the models are moreheuristic or rulelike. MBR models typically cap-ture causal relationships to diagnose problemsor predict situation outcomes.

The first system to combine CBR with MBRwas CASEY, which diagnosed heart failures(Koton 1988). CASEY was a primarily CBR sys-tem that interfaced with a previously existingMBR heart failure program. It used knowledge

Articles

SPRING 2002 75

coloring plastics, and SOPHIST (Aarts and Rousu1997), which plans bioprocess recipes, such asthose for brewing beer.

Genetic algorithms: Genetic algorithms aretechniques patterned after the evolutionaryprocess of natural selection. They learn goodproblem solutions by beginning with arbitrarysolutions and creating future generations ofbetter solutions by mutating and propagatingthe strongest solutions from preceding genera-tions. Maher (1998) has combined geneticalgorithms with CBR in an architectural designdomain. In architectural design, it is not suffi-cient to recall a past blueprint that would meetcurrent needs. Rather, some new synthesis thatdraws on past designs must be achieved. Todate, methods that have attempted to addressthis issue have been time consuming andknowledge intensive. Genetic algorithms pro-vide an efficient means of combining pieces ofpast designs. Genetic algorithms are used tocombine pieces of past building designs intonew composite designs. Design cases are repre-sented as genes, which are manipulated usingthe traditional genetic algorithm techniques ofcrossover and mutation. Many building de-signs are quickly generated and tested, allow-ing satisfactory designs to be identified andmaintained.

Information retrieval: Information re-trieval applies primarily statistical techniquesto the problem of locating appropriate textdocuments in large, unstructured collections,such as newspaper archives, legal libraries, and

representing a physiological model of the heartto match new cases to old ones and derive newdiagnoses from old ones. When CASEY couldnot find a close enough match in its case base,it invoked the original MBR system to solve theproblem. Although the MBR system was accu-rate and reliable, CBR enabled CASEY to improvesystem efficiency.

A more recent CBR-MBR hybrid is CARMA, anadvisory system that helps ranchers managegrasshopper infestations (Branting 1998). CAR-MA predicts forage loss and provides a cost-ben-efit analysis for the treatment options in a giv-en situation. It incorporates numeric modelsdeveloped by entomologists as well as specificcases of past infestations. In this domain, theavailable numeric models are not completeenough to accurately predict the effects ofalternate treatment options. Further, there arenot enough documented cases for CBR aloneto provide adequate assistance to ranchers. InCARMA, CBR is used to select a similar case, andthe model assists in adapting that case’s predic-tion by simulating the effects of different treat-ment options. The CBR-MBR integrationimproves solution accuracy over that which ispossible using either single approach. Theapproach used in CARMA, called approximate–model-based adaptation, enables prediction andcontrol planning in other domains where indi-vidual knowledge sources are insufficient forthese tasks. Two other systems that have usedthis approach are FORMTOOL (Cheatham andGraf 1997), which plans colorant formulas for

Articles

76 AI MAGAZINE

Table 1. Case-Based Reasoning Steps Supported by Constraint-Satisfaction Problem Solving.

System Case Representation Case Adaptation Case Acquisition

Weigel’s No Yes Yes

COMPOSER Yes Yes No

CADRE Yes Yes No

IDIOM Yes Yes No

CHARADE Yes Yes No

ADIOP Yes No No

the World Wide Web. Typically, users indicatethe type of document they deem appropriateby entering structured queries containing key-words. An information-retrieval system strivesto achieve perfect recall, or retrieval of all rele-vant documents, and perfect precision, orretrieval of only relevant documents. Thedegree of recall and precision actually achievedcan depend on the user’s skill in formulatingqueries.

SPIRE, a CBR–information-retrieval integra-tion, allows users to retrieve documents from alarge legal corpus without having to specifyformal queries (Daniels and Rissland 1997;Rissland and Daniels 1996). It uses cases oflegal documents and passages to generatequeries, which are then run by a full-textretrieval engine. SPIRE first analyzes a new prob-lem with respect to its own case base in theway standard for legal CBR systems. It thenuses the results of this analysis to drive theINQUERY information-retrieval system, whichthrough its relevancy feedback moduleretrieves documents in its standard way. TheCBR module of SPIRE analyzes the problem andselects the most relevant cases that it has in itscase base. It hands the texts of these cases toINQUERY, which automatically generates a stan-dard query composed of terms or pairs ofterms. Thus, SPIRE bridges the gap between thepurely symbolic CBR engine and the purelytext-based retrieval engine. This enabled SPIRE

to achieve excellent results, particularly whenlarge numbers of terms (100 or more) wereused. For example, in one experiment, SPIRE wastested on a problem case based on an actual taxcase (the Weissman case) (Rissland and Daniels1996) where its CBR module retrieved casescausing its information-retrieval module togenerate a query of 150 terms. SPIRE returnedseveral important home-office deduction cases,including two that were factually right onpoint in that they involved professors takinghome-office deductions, just as in the problemcase. Perhaps more importantly, SPIRE found themost important case (Soliman) on the topic inall its three versions: (1) a 1990 lower TaxCourt case; (2) a 1991 case in the Court ofAppeals; and (3) a 1993 Supreme Court case,which still is the definitive case on this topic.None of these cases were known to SPIRE, andall the Soliman cases were decided after SPIRE’scase base was created. Thus, not only was SPIRE

able to find highly relevant documents, it wasalso able to cope successfully with the “stale-ness” problem by accessing documents createdlong after its own case base.

SPIRE uses a similar approach in retrievingspecific passages from within a large docu-

ment. In this “inner loop,” SPIRE uses pastexamples of text discussing an issue (for exam-ple, the honesty of a debtor applying for bank-ruptcy relief) to generate a query to run on thedocument to locate passages where the issue isdiscussed. SPIRE allows ordinary users to runqueries at an expert level, minimizing the timethey spend reviewing retrieved documents thatthey do not deem to be appropriate or relevant.SPIRE is an example of CBr in that CBR is used(twice!) to aid information retrieval, first tofind relevant documents and then within eachdocument to find relevant passages.

The STAMPING ADVISER system is anotherCBR–information-retrieval integration, whichaids feasibility engineers in evaluating designsfor stamped metal automotive parts at Ford(Leake et al. 1999). Here, CBR and informationretrieval are integrated into the engineer’soverall work flow to support the entire designprocess. Stamped parts make up the bodies ofautomobiles and, as such, are designed to meetaesthetic, structural, functional, cost, and man-ufacturability criteria. They are initially de-signed using a computer-aided design (CAD)tool, and then they are critiqued in a feasibilityanalysis to identify any potential problemsbefore the design is finalized. Past cases ofstamped parts, along with problems encoun-tered in their design and solutions to thoseproblems, are used in the STAMPING ADVISER.These past cases are integrated with other avail-able knowledge sources, such as companyguidelines, to ensure consistent styling at a rea-sonable cost. The STAMPING ADVISER uses just-in-time retrieval; that is, it presents informationwhen it is needed rather than require the userto make explicit queries. Based on the task thefeasibility engineer is engaged in, the systemautomatically generates queries to retrieve andpresent relevant style guidelines. Thus, CBRand information retrieval are integrated intothe larger task context, providing the feasibili-ty engineer with a tool that is natural to use.

Case-based reasoning: Researchers havealso suggested integrating multiple CBR com-ponents within a single system. This approachis useful when a problem domain requires twodifferent styles of CBR, such as transformation-al and derivational CBR, or two different levelsof reasoning, such as domain-specific reason-ing and metareasoning. A system that exempli-fies this approach is DIAL, which operates in thedomain of disaster response planning (Leakeand Kinley 1998). It generates high-level plansfor reacting to disasters, such as floods andearthquakes. DIAL might be considered a CBR-RBR hybrid in that RBR is used to support sim-ilarity assessment and case adaptation. Howev-

Articles

SPRING 2002 77

equal footing. Its domain-independent archi-tecture has the following major components:(1) there are two independent co-reasoners,one case based and one rule based; (2) each rea-soner has a dedicated monitor that makesobservations on its processing, results, and par-tial results and recasts these observations in alanguage understandable by the controller;and (3) an agenda-based controller uses theobservations of the monitors to propose andselect tasks to be acted on by the individual co-reasoners.

CABARET’s system architecture allows it tocombine CBR and RBR in a highly opportunis-tic way. The system architecture, first describedin Rissland and Skalak (1991), is shown in fig-ure 5. CABARET works as follows: One of the co-reasoners works on a task; the monitor mod-ules make observations; and based on theobservations, the controller posts new tasksand selects the next one to be worked on. Onconclusion, CABARET outputs a memo generatedby filling in a stylized template.

CABARET uses four distinct sources of knowl-edge: (1) CBR knowledge, including a casebase, indexes, and similarity metrics; (2) RBRknowledge, including rules and predicates; (3)control knowledge, or domain-independentcontrol rules to propose and rank tasks basedon observations made by the monitors; and (4)general domain knowledge, especially hierar-chies available to all modules.

CABARET’s control rules guide its problemsolving and embody its theory of statutoryinterpretation. Examples are (1) if one mode ofreasoning fails, then switch to the other; (2)once a conclusion is reached, switch the formof reasoning to check if it holds in the othermode; (3) if all but one antecedent of a rule canbe established, then use CBR to show themissed antecedent can be established usingcases; and (4) if all but one antecedent of a rulecan be established, then use CBR to show themissed antecedent is not necessary. The lasttwo examples are “near-miss” heuristics. Notonly are they particularly useful in rule inter-pretation and argument in legal reasoning(Skalak and Rissland 1992), they could also beused in any mixed case-and-rule domain inwhich the rules are not logically ironclad.

Note that interpretive CBR is not confined toprofound problems of law and philosophy. It isubiquitous in everyday life. For example, mostparents eventually face an instance of decidingwhether or not to give their high school–agedson or daughter the keys to the family car todrive to an event (a rock concert, the seniorprom) many miles away from home and likelyto involve late night driving. Most children

er, its strength lies in integrating different CBRcomponents to monitor, capture, and replay itsreasoning processes. DIAL’s controlling processis a transformational CBR module that gener-ates new disaster response plans by remember-ing and adapting old ones. It is supported byderivational CBR modules that assist with sim-ilarity assessment, case retrieval, and case adap-tation.

Tasks Benefiting from Case-Based–Reasoning Integration

Among tasks that have benefited from CBRintegration are interpretation and argumenta-tion, design and synthesis, planning, and themanagement of long-term medical conditions.

Interpretation and ArgumentationInterpretive CBR is frequently used in disci-plines such as ethics, public policy, and law,where new fact situations are interpreted inlight of past ones. In these disciplines, con-cepts are often open textured in that they can-not be defined by universally valid necessaryand sufficient conditions but, rather, have apenumbra of borderline cases whose interpre-tation is arguable (that is, they could reason-ably be classified either way, as positive or neg-ative instances of the concept), and expertsoften disagree and argue about their interpreta-tion. Appellate law typically deals with diffi-cult, penumbral cases. In fact, CBR is the pri-mary mode of reasoning underlyingAnglo-American law. Such precedent-basedlegal systems are based on the doctrine of staredecisis that mandates that similar cases shouldbe decided similarly.

Statutory legal domains abound in interpre-tive CBR and RBR. A good example of such adomain is tax law, which is based on statutoryrules and regulations but whose rules are notironclad and whose ingredient terms—eventhe term income—are not well defined. Despitethese shortcomings, they must nevertheless beapplied to myriad situations. To inform theirapplication, all involved—taxpayers, InternalRevenue Service administrators, accountants,lawyers, tax courts—use cases. For example,the so-called home-office deduction rulerequires that a taxpayer use his/her homeoffice on a “regular basis” (Section 280A(c)(1)of the tax code), but nowhere is there a defini-tion of what constitutes such usage. The bestguide is to look at how this term has beeninterpreted in past cases. The CABARET system(Rissland and Skalak 1991), one of the firstCBR-RBR hybrids, operates in this domain.

CABARET’s control regime gives CBR and RBR

Among tasksthat have

benefited fromCBR

integration areinterpretationand argumen-tation, designand synthesis,planning, and

the manage-ment of long-term medical

conditions.

Articles

78 AI MAGAZINE

face a situation of making the “case” for suchpermission. Parents and child might arguetheir positions by considering how suchrequests were handled in past cases, for exam-ple, those involving older siblings, classmates,and neighbors. They would no doubt delveinto such issues as driving experience, pastbehavior in situations requiring responsibleaction, the perceived hazards of the trip, likely

weather and road conditions, and the commu-nity standards of the child’s and the parents’peer groups.

The end result of this type of CBR is not onlyan interpretation of the new case but also arationale justifying it. A mere yes or no wouldnot suffice in the previous example (especiallyif the “petitioner” loses). Case-based jus-tification involves comparing and contrasting

Articles

SPRING 2002 79

Control Heuristics

Agenda

Goals Satisfied

Derived Facts

Relevant Cases

RBR Monitor CBR Monitor

Rule-Based

Rules

Case-BasedReasonerReasoner

Cases

Indices

Metrics

...

...

...

...

...

...

Argument

Explanation

Supporting Cases, Rules, Facts

User Input

Controller

System Output—Analysis of Case

Top Level Purpose (argument, explanation)

Facts of Current Case

Point of View (pro, con)

Figure 5. CABARET’s Architecture.

must constrain intake of calories, fat, sodium,calcium, protein, cholesterol, or other nutri-ents. In designing menus manually, nutrition-ists naturally use both CBR and RBR. CAMPER

was built by combining the best features ofindependent CBR and RBR nutritional menu-planning systems. The CBR component stores,retrieves, and adapts daily menus, and the RBRcomponent relies on a process of generate,test, and repair, using menu patterns, anontology of foods, and common-sense knowl-edge of the ways in which foods can be com-bined. CBR and RBR complement each otherin CAMPER. CBR contributes an initial menuthat meets design constraints by capitalizingon food combinations that have proven satis-factory in the past. RBR allows analysis ofalternatives, so that innovation beyond thetried and true becomes possible. After an ini-tial menu is proposed through CBR, the userinteracts with an RBR what-if analysis moduleto explore creative alternatives. The user canmodify the menu, guided by rules, and thensave significantly altered menus in the casebase for future use.

Another CBR-RBR hybrid, SAXEX, integratesbackground musical knowledge into a primar-ily CBR system for generating expressive musi-cal performances (Arcos, Lopez de Mantaras,and Serra 1998). This application is of practicaluse in music performance and synthesisbecause musical scores contain informationabout the pitch and duration of each note, butthey do not help with the expressive parame-ters of rubato, vibrato, articulation, and attack.These parameters are ordinarily inferred andincorporated into a performance by a humanperformer. SAXEX operates in the context oftenor saxophone interpretation of standardjazz ballads. SAXEX infers a set of expressivetransformations to apply to an inexpressivemusical input phrase and then outputs anexpressive performance of the same phrase.Spectral modeling and synthesis techniquesare used to extract high-level parameters fromexpressive musical performances. The expres-sive performances are stored in SAXEX’s casebase. A case in SAXEX consists of a musical scoreplus the extracted parameters that allow it tobe played expressively. Background musicalknowledge is used for identifying cases that aresimilar to an input problem and transforminga new inexpressive piece based on features of astored expressive piece. Experiments run onSAXEX showed results comparable to humanperformance. Researchers plan to extend SAXEX

to make it a practical tool for musicians.

the new case with relevant precedents; creatinganalogies between the new case and the sup-porting precedents; and breaking analogieswith, or “distinguishing away,” contrary cases.

Design and SynthesisDesign is the process of fully describing a newartifact that will, when produced, satisfy a giv-en set of requirements. In case-based design,past designs or design experiences are used inthe creation of new designs. The task of designis one in which integrating other approacheswith CBR has proven fruitful.

FABEL is a CBR-MBR-RBR system for architec-tural design (Gebhardt et al. 1997). A buildingin FABEL contains tens of thousands of designobjects. It comprises several cases, each con-taining from two to a few hundred designobjects. Some design objects are concrete, suchas walls or pipes, and others are more abstract,such as intended use or climate. The design ofone building encompasses many problems,each of which can be solved by differentmeans. FABEL provides architects and civil engi-neers with an integrated design environmentcalled a virtual construction site. It includes acomputer-aided architectural design (CAAD)tool, an object-oriented distributed database,and several problem solvers. Some problemsolvers are case based, but others use modelsfor building subsystems, and one uses rulesthat have been induced from the cases in thesystem. Gebhardt et al. maintain that the com-plexity of this real-world problem makes ituntenable to use single design cases or a singleproblem-solving strategy.

CAMPER is a design system that combinesCBR with RBR (Marling and Sterling 1998). Itdesigns a daily menu for an individual inaccordance with nutrition guidelines, personalpreferences, and aesthetic criteria. Planningnutritional menus is a task researchers havetried to automate since the early 1960s (Balint-fy 1964). The first computer-based menu plan-ner used linear programming to optimize amenu for nutritional adequacy, cost, and con-sumer satisfaction. The menus it generatedmet its first two goals but were not consideredappetizing by people. Subsequent attemptsalso failed, in part because rules for evaluatingmenus were insufficient for generating satis-factory menus in the first place. JULIA, the CBR-CSP system that planned menus for dinnerparties, showed that past menus could providea basis for planning appetizing new ones (Hin-richs 1992). However, it was a self-proclaimed“party animal” and did not try to incorporatenutrition criteria. CAMPER is used by a qualifiednutritionist in consultation with a client who

Articles

80 AI MAGAZINE

PlanningA planning task consists of constructing asequence of actions to achieve a specified set ofgoals when starting from an initial situation. Aclassical generative planning process involvessearching a space of possible actions to obtainthe sequence of actions solving the problem.CBR techniques have been shown to be effec-tive for two main reasons: First, frequently,there is no complete domain theory that couldbe used to generate plans from scratch. In thissituation, cases serve as additional knowledgeabout the domain. Second, in domains wherea complete domain theory is available, gener-ating plans from scratch can be computation-ally expensive. CBR can be used as a source ofmetaknowledge indicating how to use domaintheory to solve a given problem.

Most CBR approaches to planning that donot assume the existence of a complete domaintheory use transformational analogy for planadaptation, although derivational analogy canalso be used. In transformational analogy, casesare transformed to obtain a solution plan, as inDIAL, the disaster response planner, which usestransformational analogy as its control processand derivational analogy for support (Leakeand Kinley 1998). In derivational analogy, casescontain annotated derivational traces, whichare followed when solving a new problem.Derivational analogy approaches are most fre-quently used in connection with first-principlesplanning systems.

CBR has been combined with two types offirst-principles planning: (1) STRIPS-style plan-ning and (2) hierarchical planning. STRIPS-styleplanning originated in the Stanford ResearchInstitute Problem Solver (STRIPS) Project of theearly 1970s. STRIPS was a general problem solverthat gave instructions to SHAKEY the robot as itnavigated its way through different rooms,moving objects from one room to another(Fikes and Nilsson 1971). In STRIPS-style plan-ning, the assumption is made that changesonly occur in the world if indicated by theoperators in the domain definition. Theseoperators are used to generate the ordered stepsof a plan that, when executed, will achieve theintended goal. In hierarchical planning, planshave several levels of abstraction. Complextasks are decomposed into simpler ones. Plansat the most concrete level consist of sequencesof STRIPS-style operators, but planning begins athigher levels of abstraction. Hierarchical plan-ning has been proven to be strictly moreexpressive than one-level STRIPS-style planning(Erol, Nau, and Hendler 1994).

The first system to combine CBR with STRIPS-style planning was PRODIGY/ANALOGY (Veloso

1994). PRODIGY/ANALOGY is a generative case-based planner. It is a total-order planning sys-tem, which means that the steps appear in thefinished plan in the same order in which theyare generated. PRODIGY/ANALOGY is domain inde-pendent, but it has been validated in a logisticstransportation domain in which packages aremoved among different locations by trucks andairplanes. Cases in PRODIGY/ANALOGY are used toguide the search process of the integrated plan-ning system. Because the search space is of expo-nential size on the length of the plans, planningby first principles can be inefficient. The integra-tion results in a significant improvement interms of the time needed to generate plans. Twoother systems that integrate CBR with STRIPS-styleplanning to improve overall efficiency are DER-SNLP (Ihrig and Kambhampati 1994) andCAPLAN/CBC (Munoz-Avila and Weberskirch1996). Unlike PRODIGY/ANALOGY, these two sys-tems use partial-order planning, in which theorder of steps in the plan is decoupled from theorder in which the steps are generated. Both sys-tems are domain independent, but DERSNLP wastested on PRODIGY/ANALOGY’s logistics transporta-tion domain, and CAPLAN/CBC was tested in aprocess-planning domain for manufacturingmechanical workpieces.

CBR has also been integrated into two hierar-chical planning systems: (1) the joint maritimecrisis action planning (JMCAP) system (des-Jardins, Francis, and Wolverton 1998) and thehierarchical interactive case-based architecturefor planning (HICAP) system (Munoz-Avila et al.1999). JMCAP and HICAP support military oper-ations. JMCAP extends an existing hierarchicaltask network (HTN) by integrating past maritimecrisis action-planning experience. The rationalefor this extension is that it more closely modelsthe processes used by human planners. HICAPwas tested on planning noncombatant evacua-tion operations (NEOs). NEOs are conductedwhen nonmilitary personnel must be evacuatedfrom dangerous situations and moved to safehavens. Here, there is no complete domain the-ory to use in generating plans. Cases provideknowledge about previous military operations,which would otherwise be unavailable for plan-ning. Cases complement the knowledge provid-ed by doctrine, the general guidelines for mili-tary planning, and standard operationalprocedures, the documented methods forachieving military objectives.

Management of Long-Term Medical ConditionsThe medical domain has provided fertileground for CBR integration research ever sincethe CBR-MBR system CASEY (Koton 1988),

Articles

SPRING 2002 81

table 2. Much can be learned from these exam-ples. However, much remains to be learnedbefore system developers will be able to selectjust the right integration strategy in advance ofsystem implementation. Thorough evaluationsof integrated systems are needed to identify thecontributions of each paradigm to the whole.New studies are needed that do not stop withwhat “works” for a particular application butthat explore beyond working solutions todevelop an understanding that could lead toeven better solutions.

Open issues in CBR integration include find-ing the best system architectures and knowl-edge representations for hybrid systems. Onekey issue in multimodal reasoning, whether byperson or program, is how the different modesof reasoning are integrated. If we consider CBR-RBR integrations, we find that the modes ofreasoning with rules and cases are quite differ-ent—logical reasoning for rules and analogicalreasoning for cases—and thus can be per-formed by separate modules that must, howev-er, be coordinated. For example, is CBR only tobe used to rescue RBR from difficulties or pro-vide a sanity check, or is the processing moreintimately intertwined? How can CBR enhancethe other mode, and vice versa? Many hybridsuse CBr to narrow the search space for the oth-er mode by seeding or localizing problem solv-ing to “neighborhoods” around the most rele-vant cases. Some integrations use failures inthe non-CBR mode to trigger acquisition ofcases. In most of the applications discussed inthis article, the integration of CBR with anoth-er mode of reasoning is “straight line” in thesense that first the program performs one andthen the other. In other systems, such asCABARET and CARE-PARTNER, processing is moreintertwined.

At the 1998 American Association for Artifi-cial Intelligence Workshop on CBR Integra-tions, Aha and Daniels (1998) trichotomizedprevailing system architectures as follows: (1)master-slave: other reasoning methods supportthe CBR component; (2) slave-master: the CBRcomponent supports other reasoning methods;(3) collaborating: there are less stratified collab-orations. GYMEL, the system that harmonizesmelodies, is an example of the master-slavearchitecture. CBR is the dominant mode, butRBR assists when insufficient cases are availableto suggest a harmonization. ADIOP, which diag-noses interoperability problems in ATM net-works, exemplifies the slave-master architec-ture. Here, CSP is the dominant paradigm, butCBR is used to compensate when CSP modelsare incomplete. CARE-PARTNER uses a collaborat-ing architecture, intertwining CBR, RBR, and

which diagnosed heart failures. Although earlyAI in medicine systems focused on diagnosis, amore recent research emphasis is on systemsthat support the treatment of chronic, or long-term, conditions, which cannot be cured butnevertheless must be managed in the bestinterests of the patient. Both problem-solvingand interpretive CBR come into play, as do oth-er reasoning modalities.

CARE-PARTNER assists clinicians responsible forthe long-term follow-up care of cancer patientswho have undergone bone-marrow transplants(Bichindaritz, Kansu, and Sullivan 1998). Itintegrates CBR, RBR, and information retrievalto support evidence-based medical practice.Cases are used to show how problems weresolved for past patients. Rules are derived fromstandard practice guidelines and special prac-tice pathways developed by experts. Informa-tion retrieval is used to provide relevant docu-ments from the medical literature that cannotreadily be translated into rules. CARE-PARTNER

stores cases, rules, and documents in a unifiedknowledge base. No reasoning modality isdominant; rather, reasoning steps from differ-ent modalities are merged based on the prob-lem at hand.

The Auguste Project is a current effort to pro-vide decision support for planning the ongo-ing care of Alzheimer’s disease patients (Mar-ling and Whitehouse 2001). Formulated as astudy in CBR integration, researchers are focus-ing first on those reasoning modalities natural-ly used in this task by physicians, nurses, andsocial workers, including CBR and RBR. Theynext plan to explore whether less intuitiveapproaches, such as Bayesian reasoning, mightaugment the natural reasoning processes. Geri-atrics practitioners rely heavily on case studiesfor training, making suggestions for treatment,sharing best practices, and reporting clinicalfindings. They also rely on guidelines, or rules,to determine which interventions to prescribeand to tailor prescribed interventions to partic-ular patients. Because the symptoms of Alz-heimer’s disease vary greatly from patient topatient, and within the same patient over time,there is no critical pathway, or overall set ofrules, that can be applied to all patients.Although the study is still ongoing, CBR hasalready proven useful for classifying patientsinto subcategories, within which small andmanageable rule sets can apply.

Open Issues and Future WorkIn a case-based manner, many examples, orcases, of CBR integration have been presented.These example systems are summarized in

… muchremains to belearned before

systemdevelopers

will be able toselect just the

rightintegrationstrategy inadvance of

systemimplemen-

tation.

Articles

82 AI MAGAZINE

information retrieval to assist with the long-term follow-up care of cancer patients. Partici-pants at the AAAI-98 Workshop on CBR Inte-grations categorized their own systems,reporting 11 master-slave integrations, 6 slave-master integrations, and 12 collaborating rea-soners.

Another key issue is the representation ofknowledge. Different modalities have distinctways of structuring knowledge. RBR relies onrules, which represent knowledge by express-

ing relationships, typically in the form ofantecedents and consequents. Each rule con-tains a fairly small unit of knowledge, so mul-tiple rules are chained together during infer-encing. CSP represents knowledge asconstraint graphs in which nodes are variables,labels on nodes are values, and edges are con-straints. A single graph contains a problemdescription. MBR uses mathematical equationsand relationships. Genetic algorithms repre-sent knowledge as genes, which are bit strings

Articles

SPRING 2002 83

Table 2. Summary of Case-Based Reasoning Integrations.

System Domain1 Task CBR PlusADIOP ATM networks Diagnosis CSPANAPRON Speech Synthesis RBRAUGUSTE Project Medicine Ongoing treatment RBRCABARET Law Argumentation, interpretation RBRCADRE Architecture Design CSPCADSYN Architecture Design CSPCAMPER Menu planning Design RBRCAPLAN/CBC Manufacturing Planning STRIPS planningCARE-PARTNER Medicine Ongoing treatment RBR, InformationretrievalCARMA Entomology Prediction, planning MBRCASEY Medicine Diagnosis MBRCHARADE Forest fires Planning CSPCOMPOSER Engineering Configuration CSPDERSNLP Transportation Planning STRIPS planningDIAL Disaster response Planning CBR componentsFABEL Architecture Design MBR, RBRFORMTOOL Plastic colorants Planning MBRGREBE Law Argumentation, interpretation RBRGYMEL Music Synthesis RBRHICAP Military Planning Hierarchical planningIDIOM Architecture Design CSPIKBALS Law Argumentation, interpretation RBR, InformationretrievalJMCAP Military Planning Hierarchical planningJULIA Menu planning Design CSPMaher’s Architecture Design Genetic algorithmsPRODIGY/ANALOGY Transportation Planning STRIPS planningSAXEX Music Synthesis RBRSOPHIST Bioprocess recipes Planning MBRSPIRE Law Document retrieval Information retrievalSTAMPING ADVISER Engineering Design Information retrievalWeigel’s Products for sale Configuration CSP

1. For domain-independent systems, the primary test domain is listed.

CSP = constraint-satisfaction problem.RBR = rule-based reasoning.MBR = model-based reasoning.

New studies are needed to clarify the advan-tages and disadvantages of the various systemarchitectures and knowledge representationsthat can be used. Future efforts in this areamight well lead to discovery of new synergies,greater methodological support for buildingpractical integrated systems, and a betterunderstanding of the ways in which peopleand machines reason.

AcknowledgmentsThis survey article grew out of discussions thatwere initiated at the AAAI-98 Workshop onCBR Integrations and continued by the coau-thors. The coauthors are grateful to the work-shop participants for their insights and contri-butions.

ReferencesAamodt, A., and Plaza, E. 1994. Case-Based Reason-ing: Foundational Issues, Methodological Variations,and System Approaches. AI Communications 7(I):39–59.

Aarts, R., and Rousu, J. 1997. Qualitative Knowledgeto Support Reasoning about Cases. In Proceedings ofthe Second International Conference on Case-Based Rea-soning, 489–498. Berlin: Springer.

Aha, D., and Daniels, J. J., eds. 1998. Case-Based Rea-soning Integrations: Papers from the 1998 Workshop.Menlo Park, Calif.: AAAI Press.

Arcos, J. L.; López de Mántaras, R.; and Serra, X.1998. Integrating Background Musical Knowledge ina CBR System for Generating Expressive Musical Per-formances. In Case-Based Reasoning Integrations:Papers from the 1998 Workshop, 12–16. Menlo Park,Calif.: AAAI Press.

Ashley, K. D. 1990. Modeling Legal Argument: Reason-ing with Cases and Hypotheticals. Cambridge, Mass.:MIT Press.

Avesani, P.; Perini, A.; and Ricci, F. 1993. CombiningCBR and Constraint Reasoning in Planning ForestFire Fighting. In Proceedings of the First EuropeanWorkshop on Case-Based Reasoning, 325–328. Berlin:Springer.

Balintfy, J. L. 1964. Menu Planning by Computer.Communications of the ACM 7(4): 255–259.

Barletta, R. 1994. Comments from Address Given atthe Second European Conference on Case-Based Rea-soning (ECCBR-94), 7–10 November, Chantilly,France.

Bichindaritz, I.; Kansu, E.; and Sullivan, K. M. 1998.Integrating Case-Based Reasoning, Rule-Based Rea-soning, and Intelligent Information Retrieval forMedical Problem Solving. In Case-Based ReasoningIntegrations: Papers from the 1998 Workshop, 22–27.Menlo Park, Calif.: AAAI Press.

Branting, L. K. 1998. Integrating Cases and Modelsthrough Approximate Model-Based Adaptation. InMultimodal Reasoning: Papers from the 1998 AAAISpring Symposium, 1–5. Menlo Park, Calif.: AAAIPress.

Branting, L. K. 1991. Building Explanations from

in which each part of the string stands for anattribute-value pair. Information retrieval doesnot structure knowledge at all; rather, it is usedto find information that has not been struc-tured.

There is considerable flexibility in represent-ing a case in a CBR system. It is important tocapture a description of a past problem, thepast solution to the problem, and the outcomeof applying this solution to the problem. Howthese are captured varies greatly from system tosystem. Integrations in which cases are directlyrepresented by constraint graphs or genes takeadvantage of this flexibility. However, mostintegrations to date maintain different knowl-edge representations for different reasoningmodalities. What are the advantages anddisadvantages of each representation? Couldknowledge representations themselves be inte-grated to develop useful composite knowledgestructures? More research is needed to answerthese questions.

Summary and ConclusionsOver 10 years ago, Rissland and Skalak (1989)wrote, “At this point, researchers have onlyrecently begun to write about the integrationof CBR with other reasoning paradigms. Wefeel that such mixed-paradigm approaches arenatural and shed light on both the cognitiveskills involved in such reasoning and on ques-tions of architecture and control of their com-putational models” (p. 526). Today, researchershave uncovered many synergistic ways to com-bine CBR with other modes of reasoning. CBRhas fruitfully been combined with RBR, CSPsolving, MBR, genetic algorithms, and infor-mation retrieval. CBR integrations have sup-ported a wide range of tasks, including inter-pretation and argumentation, design andsynthesis, planning, and management of long-term medical conditions. Many useful syner-gies have emerged as different reasoning strate-gies extend and complement each other.Integrated systems have enabled more accuratemodeling of domain knowledge, compensa-tion for incomplete domain models and rulebases, compensation for small case bases, sim-plification of knowledge acquisition, improvedsolution quality, improved system efficiency,leveraging of past experiences, and compensa-tion for shortcomings inherent in individualreasoning strategies.

More research is needed because open issuesremain to be resolved. Thorough system evalu-ations are needed to pinpoint the specificstrengths and weaknesses of different reason-ing strategies for performing different tasks.

Articles

84 AI MAGAZINE

Rules and Structured Cases. International Journal ofMan-Machine Studies 34(6): 797–837.

Cheatham, W., and Graf, J. 1997. Case-Based Reason-ing in Color Matching. In Proceedings of the SecondInternational Conference on Case-Based Reasoning,1–12. Berlin: Springer.

Daniels, J. J., and Rissland, E. L. 1997. What You SawIs What You Want: Using Cases to Seed InformationRetrieval. In Proceedings of the Second InternationalConference on Case-Based Reasoning, 325–336. Berlin:Springer.

desJardins, M.; Francis, A.; and Wolverton, M. 1998.Hybrid Planning: An Approach to Integrating Gener-ative and Case-Based Planning. In Case-Based Reason-ing Integrations: Papers from the 1998 Workshop,45–49. Menlo Park, Calif.: AAAI Press.

Erol, K.; Nau, D.; and Hendler, J. 1994. HTN Plan-ning: Complexity and Expressivity. In Proceedings ofthe Twelfth National Conference on Artificial Intelli-gence, 1123–1128. Menlo Park, Calif.: AmericanAssociation for Artificial Intelligence.

Faltings, B. 1996. Working Group in Model-BasedDesign and Reasoning. Part II: Design. AI Communi-cations 9(2): 65–73.

Fikes, R. E., and Nilsson, N. J. 1971. STRIPS: A NewApproach to the Application of Theorem Proving inProblem Solving. Artificial Intelligence 5(2): 189–208.

Freuder, E., and Mackworth, A. 1992. Constraint-Based Reasoning. Artificial Intelligence (Special Issue)58.

Gebhardt, F.; Voss, A.; Gräther, W.; and Schmidt-Belz,B. 1997. Reasoning with Complex Cases. Boston, Mass.:Kluwer Academic.

Golding, A. R., and Rosenbloom, P. S. 1991. Improv-ing Rule-Based Systems through Case-Based Reason-ing. In Proceedings of the Ninth National Confer-ence on Artificial Intelligence, 22–27. Menlo Park,Calif.: American Association for Artificial Intelli-gence.

Hammond, K. J. 1989. Case-Based Planning: ViewingPlanning as a Memory Task. San Diego, Calif.: Acade-mic.

Hammond, K. J. 1986. CHEF: A Model for Case-BasedPlanning. In Proceedings of the Fifth NationalConference on Artificial Intelligence, 267–271. Men-lo Park, Calif.: American Association for ArtificialIntelligence.

Hart, H. L. A. 1958. Positivism and the Separation ofLaw and Morals. Harvard Law Review 71(4): 593–629.Reprinted in Hart, H. L. A. 1983. Essays in Jurispru-dence and Philosophy. Oxford: Clarendon Press.

Hayes-Roth, F.; Waterman, D. A.; and Lenat, D. B.,eds. 1983. Building Expert Systems. Reading, Mass.:Addison-Wesley.

Hinrichs, T. R. 1992. Problem Solving in Open Worlds:A Case Study in Design. Hillsdale, N. J.: Lawrence Erl-baum.

Hua, K., and Faltings, B. 1993. Exploring Case-BasedBuilding Design—CADRE. AI in Engineering Design,Analysis and Manufacturing 7(2): 135–143.

Ihrig, L. H., and Kambhampati, S. 1994. DerivationReplay for Partial-Order Planning. In Proceedings of

the Twelfth National Conference on Artificial Intelli-gence, 992–997. Menlo Park, Calif.: American Asso-ciation for Artificial Intelligence.

Kolodner, J. 1993. Case-Based Reasoning. San Francis-co, Calif.: Morgan Kaufmann.

Kolodner, J. L., and Leake, D. B. 1996. A TutorialIntroduction to Case-Based Reasoning. In Case-BasedReasoning: Experiences, Lessons, and Future Directions,ed. D. B. Leake, 31–66. Menlo Park, Calif.: AAAIPress.

Koton, P. 1988. Reasoning about Evidence in CausalExplanations. In Proceedings of the Seventh Nation-al Conference on Artificial Intelligence, 256–261.Menlo Park, Calif.: American Association for Artifi-cial Intelligence.

Kumar, V. 1992. Algorithms for Constraint-Satisfac-tion Problems: A Survey. AI Magazine 13(1): 32–44.

Leake, D. B., and Kinley, A. 1998. Integrating CBRComponents within a Case-Based Planner. In Case-Based Reasoning Integrations: Papers from the 1998Workshop, 80–84. Menlo Park, Calif.: AAAI Press.

Leake, D. B.; Birnbaum, L.; Hammond, K.; Marlow,C.; and Yang, H. 1999. Integrating InformationResources: A Case Study of Engineering Design Sup-port. In Proceedings of the Third International Confer-ence on Case-Based Reasoning, 482–496. Berlin:Springer.

Maher, M., and Zhang, D. 1991. CADSYN: Using Caseand Decomposition Knowledge for Design Synthesis.In Artificial Intelligence in Design, ed. J. S. Gero,137–150. Oxford, U.K.: Butterworth-Heineman.

Maher, M. L. 1998. CBR as a Framework for Design:Augmenting CBR with Other AI Techniques. In Case-Based Reasoning Integrations: Papers from the 1998Workshop, 96–101. Menlo Park, Calif.: AAAI Press.

Marling, C., and Sterling, L. 1998. Integrating CBRand RBR for Nutritional Menu Design. In Case-BasedReasoning Integrations: Papers from the 1998 Workshop,102–107. Menlo Park, Calif.: AAAI Press.

Marling, C. R., and Whitehouse, P. J. 2001. TheAUGUSTE Project: Decision Support for Patient Care. InProceedings of the International ICSC-NAISO Con-gress on Computational Intelligence: Methods andApplications, 65–70. Canada/The Netherlands: ICSCAcademic.

Muñoz-Avila, H., and Weberskirch, F. 1996. Planningfor Manufacturing Workpieces by Storing, Indexingand Replaying Planning Decisions. In Proceedings ofthe International Conference on AI Planning Systems.Menlo Park, Calif.: AAAI Press.

Muñoz-Avila, H.; McFarlane, D.; Aha, D. W.; Ballas,J.; Breslow, L. A.; and Nau, D. 1999. Using Guidelinesto Constrain Interactive Case-Based HTN Planning.In Proceedings of the Third International Conference onCase-Based Reasoning, 288–302. Berlin: Springer.

Puget, J. F., and Leconte, M. 1995. Beyond the GlassBox: Constraints as Objects. In Proceedings of theInternational Logic Programming Symposium, 513–527.Cambridge, Mass.: MIT Press.

Purvis, L. 1998. Synergy and Commonality in Case-Based and Constraint-Based Reasoning. In Multi-modal Reasoning: Papers from the 1998 AAAI SpringSymposium, 80–84. Menlo Park, Calif.: AAAI Press.

Articles

SPRING 2002 85

Edwina Rissland is pro-fessor of computer sci-ence at the University ofMassachusetts at Am-herst. For many years,she was also a lecturer onlaw at the Harvard LawSchool, where she taughta course on her specialty

of AI and law. She received her Sc.B. fromBrown University and her Ph.D. in mathe-matics from the Massachusetts Institute ofTechnology. She has long been interestedin how people and machines reason withexamples and cases and was one of thefounders of case-based reasoning. Her e-mail address is rissland@cs.umass. edu.

Hector Muñoz-Avila is an assistant profes-sor in the Computer Science and Engineer-ing Department at Lehigh University. Hereceived his B.S. and M.S. in computer sci-ence and B.S. in mathematics from the Uni-versity of the Andes in Columbia and hisdoctorate from the University of Kaiser-slautern in Germany. Previously, he was aresearcher at the University of Maryland atCollege Park and at the Navy Center forApplied Research in Artificial Intelligence.His e-mail address is munoz@cse.lehigh.edu.

David Aha leads the Intelligent DecisionAids Group at the Navy Center for AppliedResearch in Artificial Intelligence, a branchof the Naval Research Laboratory. Hisresearch interests currently focus on mixed-initiative planning and intelligent lessonslearned systems. Aha is an editor for theMachine Learning journal, is an editorialboard member for Applied Intelligence, andwas program cochair for the 2001 Interna-tional Conference on Case-Based Reason-ing. His e-mail address is aha@aic.navy.mil.

Constraint Satisfaction with Case-BasedReasoning. Paper presented at the NinthInternational Workshop on Principles ofDiagnosis (DX-98), 24–27 May, Cape Cod,Massachusetts.

Sqalli, M. H.; Purvis, L.; and Freuder, E. C.1999. Survey of Applications IntegratingConstraint Satisfaction and Case-BasedReasoning. Paper presented at PACLP99:The First International Conference andExhibition on the Practical Application ofConstraint Technologies and Logic Pro-gramming, 19–21 April, London, UnitedKingdom.

Veloso, M. 1994. Planning and Learning byAnalogical Reasoning. Berlin: Springer.

Wallace, M. 1996. Practical Applications ofConstraint Programming. Constraints 1(1):139–168.

Weigel, R., and Faltings, B. 1998. Inter-changeability for Case Adaptation in Con-figuration Problems. In Multimodal Reason-ing: Papers from the 1998 AAAI SpringSymposium, 69–73. Menlo Park, Calif.: AAAIPress.

Zeleznikow, J.; Vossos, G.; and Hunter, D.1994. The IKBALS Project: Multimodal Rea-soning in Legal Knowledge-Based Systems.Artificial Intelligence and Law 2(3):169–203.

Cynthia Marling is anassistant professor in theSchool of Electrical Engi-neering and ComputerScience at Ohio Univer-sity. She received herM.S. and Ph.D. in com-puter science from CaseWestern Reserve Univer-

sity. Between degrees, she worked as aknowledge engineer and project leader forGeneral Electric. Her research interests arecase-based reasoning, multimodal reason-ing, robotic soccer, and AI in medicine. Here-mail address is marling@ohio.edu.

Mohammed Sqalli is asenior automation test-ing specialist at the Tele-com Innovation Centerin Siemens Canada Lim-ited. Sqalli received hisM.S. in computer sciencefrom the University ofNew Hampshire in 1996

and his Master of Engineering in computerscience from the Mahammadia Engineer-ing School, Rabat, Morocco, in 1992. He iscurrently a Ph.D. candidate in SystemsDesign Engineering at the University ofNew Hampshire. His dissertation researchinvolves integrating constraint satisfactionand case-based reasoning. His e-mailaddress is msqalli@cs.unh.edu.

Purvis, L., and Pu, P. 1996. An Approach toCase Combination. Paper presented at theWorkshop on Adaptation in Case-BasedReasoning, Twelfth European Conferenceon Artificial Intelligence, 11–16 August,Budapest, Hungary.

Riesbeck, C. K., and Schank, R. C. 1989.Inside Case-Based Reasoning. Hillsdale, N.J.:Lawrence Erlbaum.

Rissland, E. L. 1980. Example Generation.Paper presented at the Third National Con-ference of the Canadian Society for Com-putational Studies of Intelligence, May, Vic-toria, British, Columbia.

Rissland, E. L., and Daniels, J. J. 1996. TheSynergistic Application of CBR to IR. Artifi-cial Intelligence Review (Special Issue on theUse of AI in Information Retrieval) 10(5–6):441–475.

Rissland, E. L., and Skalak, D. B. 1989.Combining Case-Based and Rule-BasedReasoning: A Heuristic Approach. In Pro-ceedings of the Eleventh InternationalJoint Conference on Artificial Intelligence,524–530. Menlo Park, Calif.: InternationalJoint Conferences on Artificial Intelligence.

Rissland, E. L., and Skalak, D. B. 1991.CABARET: Rule Interpretation in a HybridArchitecture. International Journal of Man-Machine Studies 34(6): 839–887.

Rissland, E. L., and Soloway, E. M. 1980.Overview of an Example Generation Sys-tem. In Proceedings of the First NationalConference on Artificial Intelligence,256–258. Menlo Park, Calif.: AmericanAssociation for Artificial Intelligence.

Rissland, E. L.; Valcarce, E. M.; and Ashley,K. D. 1984. Explaining and Arguing withExamples. In Proceedings of the FourthNational Conference on Artificial Intelli-gence, 288–294. Menlo Park, Calif.: Ameri-can Association for Artificial Intelligence.

Sabater, J.; Arcos, J. L.; and López de Mán-taras, R. 1998. Using Rules to Support Case-Based Reasoning for Harmonizing Mel-odies. In Multimodal Reasoning: Papers fromthe 1998 AAAI Spring Symposium, 147–151.Menlo Park, Calif.: AAAI Press.

Skalak, D. B., and Rissland, E. L. 1992.Arguments and Cases: An Inevitable Inter-twining. Artificial Intelligence and Law: AnInternational Journal 1(1): 3–48. Reprintedin Skalak, D. B., and Rissland, E. L. 1998.The Philosophy of Legal Reasoning: ScientificModels of Legal Reasoning, ed. S. Brewer,249–290. New York: Garland.

Smith, I.; Lottaz, C.; and Faltings, B. 1995.Spatial Composition Using Cases: IDIOM. InProceedings of the First International Confer-ence on Case-Based Reasoning, 88–97. Berlin:Springer.

Sqalli, M., and Freuder, E. 1998. DiagnosingInteroperability Problems by Enhancing

Articles

86 AI MAGAZINE