From: AAAI Technical Report WS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. A ComparativeUtility Analysis of Case-Based Reasoning and Control-Rule Learning Systems Anthony G. Francis, Jr. and Ashwin Ram Collegeof Computing,GeorgiaInstitute of Technology Atlanta, Georgia30332-0280 {centaur, ashwin}@ce.gatech.edu learner,2a type of systemthat attempts to improveits problem-solvingperformanceby learning search-control knowledge,called control rules, that reduce the amount of search it needs to perform by eliminating dead-end paths and selecting profitable ones. WhatMintonand others noticed about systems like PRODIGY/EBL was that the system could actually get slower after having learnedcontrol rules, rather than faster. At eachstep in the search space, a control-rule problemsolver has to matchall of its controlrules against the currentstate to determineif they shouldfire. As that library of control rules growsin size, the cost of matchingthe control rules increases to the point that it outweighsor "swamps" the savings in search they provide. 1. Introduction This side effect of learning wascalled the "utility All intelligent systemsthat learn can suffer from the problem": learning designed to improve the system’s utility problem, whichoccurs whenknowledgelearned performance ended up degrading performance instead in an attempt to improve a system’s performancede(HOLDERET AL.1990). Since Minton’sdiscoveryof this grades performanceinstead (1VIINTON1990). In this pa"swamping"utility problem in PRODIGY, researchers per, weanalyze the utility problemand examineits efhaveidentified manydifferent types of utility problems, fects in case-based reasoners and control-rule problem each manifestingitself in slightly different ways. Besolvers. Our methodologyfor this analysis couples a cause sometypes of utility problemsare affected by the functional analysis of an AI systemwith a performance hardware architecture of the system and others are analysis of the system’s algorithmic and implementa- largely independentof hardwareconcerns, wecan group tional components.Sucha computationalmodelallows the different types of utility problemsinto two rough the identification of the root causesof the utility prob- classes: architecturalutility problems,whicharise from lem, whichare combinationsof algorithmic characterisinteractions betweena system’s learning and its hardtics of an AI system(e.g., serial search of memory) with ware architecture, and search-spaceutility problems, particular "parameters"that affect its operation (e.g., which arise from interactions between learning and knowledge base size). Identifyingthe precise algorithmic problemsolvingalgorithms.Afull discussionof the difnature of a root cause allowsus to predict the threshold ferent types of utility problemsis containedin (FRANCIS conditionsunder whichit will affect a systemseverely & RAM 1994). In this paper, wewill focus on swamping, enoughto cause a utility problem. whichis the utility problemmostcommonly encountered Our analysis provides a general and theoretical in learning systems.Wewill reservethe term"the utility frameworkfor addressing this problemin systemsthat problem"for the general utility problem,and will refer havebeenstudied empirically(e.g., control-rule learning to specific versions of the utility problem, such as (CRL)systems) and in systemsfor whichlittle utility swamping,by their names. analysis has been performed(e.g., case-basedreasoning 3. A Methodologyfor Utility Analysis (CBR)systems). Weuse this frameworkto compare and CRLsystems, and find that CBRsystems are more Weproposethe use of algorithmic complexitytheory as resistant to the utility problemthan CRLsystems. a tool for the analysisof the generalutility problem.Our methodologyinvolves analyzing different types of AI 2. Analyzing the Utility Problem systemsand decomposing their cognitive architectures Theutility problemwas first detected in PRODIGY/EBL into lower-level functional units, including problem(MINTON1988). PRODIGY/EBL is control-rule solving engines and memory systems, that can be repreAbstract Theutility problem in learningsystemsoccurswhen knowledge learned in an attemptto improvea system’sperformance degradesperformance instead. We presenta methodology for the analysisof the utility problemwhichuses computational modelsof problem solvingsystems to isolatethe rootcausesof a utility problem,to detect the thresholdconditionsunder whichthe problem will arise, andto designstrategies to eliminateit. Wepresentmodels of case-based reasoningandcontrol-rulelearningsystemsandcompare themwith respect to the swamping utility problem. Ouranalysissuggeststhat CBR systemsare moreresistantto the utility problem thanCRL systems, l i Thisresearch wassupported bytheUnited StatesAirForceLaboratoryGraduate Fellowship Program andthe Georgia Instituteof Technology.Alongerversionof this paper(FRANCIS &RAM 1994)is ttp-able from 24.ps.Z ftp. cc. gatech, edu :/pub/ai/ram/git-cc- 94- 36 2 Etzioni (1992) uses the term meta-level problemsolvers for controlrule learning systems. Wehave avoided this term because of the possible confusion with metacognition, which includes systems that "knowwhat they know" (metaknowledge)and systems that reason about their own reasoningprocesses (metareasoning,or introspection). sented by formal algorithmic models. Our algorithmic approachincorporates both functional-level aspects of the computation,such as the system’scognitive architecture and its knowledgebase, and implementation-level aspects, such as the performancecharacteristics of the system’shardwarearchitecture. This multi-level analysis is crucial for the study of the utility problembecause manyutility problemsarise due to interactions between the functional level of the systemand the waythat functional computationis actually implemented.Our methodologycan be usedto identify potential utility problems as well as to designcopingstrategies to eliminatetheir effects. In this paper, wefocus on a comparative analysis of case-basedreasoningandcontrol-rule systems. matchtime and knowledgebase size; in a CBRsystem, a similar interaction exists betweencase retrieval time and caselibrary size. Wecan formally define an interaction to be a combination of a set of parameters,a module,and a set of effects. Themodulerepresents the part of the CAthat is responsible for the relationship betweenthe parameters andtheir effects. Parameters represent characteristics of the system’sknowledge base, while effects represent the performancemeasuresthat affected by the interaction. Thus,an interaction defines a function betweenlearning (changes in the knowledge base) and performance (changes in the evaluation metric), mediated by the characteristics of the algorithmiccomponent of the interaction (the module). 3.1. AI Systems and Learning Utility problemsarise whena learning modulein the Formally,wecan describe an AI systemas a triple (CA, system causes parameter changes whichinteract with KB, HA). The cognitive architecture CAspecifies a someCAcomponentto produce"side" effects that imsystemin terms of separate functional modulesthat carry pact the performancemeasuresa learning moduleis deout fixed subtasks in the system, while the knowledge signed to improve. This kind of coupling between a base KBrepresents the internal "data" that the CAuses learning moduleand an interaction is a potential root to performits computations.Thehardwarearchitecture causeof a utility problem.For a particular root cause, HAdefines the operations that a systemcan performat the calculated improvement Fc is the savings that the the implementation level, as well as their relative costs. learning moduleis designedto perform,while the actual Thecost (and hencethe utility) of an operation may cost Fa is the actual changein performance taking into different on different HA’s:for example,retrieval might account the side effects of the interaction. take longeron a serial machinethan it wouldon a paralBycomparingthe algorithmicbehaviorof the learning lel machine. module,the root cause interaction it is paired with, and the cost and savings functions that they contribute, we 3.2. Utility and the Utility Problem Utility can only be defined in terms of "performance" can computethreshoMconditions -- limiting values for the parameterchangesthat the systemcan tolerate before measuresthat judge the efficiency of a reasoner, such as the actual costs exceed the calculated improvement and executiontime, numberof states searched, storage space the systemencountersa utility problem.Eliminatingthe used, or evenquality of solution. Theseevaluationmet- general utility probleminvolves identifying the root rics measurethe costs that a system incurs during its causesof particular utility problemsthat can arise in a reasoning.Givena particular evaluationmetric, the utilsystemanddesigningcopingstrategies that preventtheir ity of a learned item can be definedas the changein exthresholdconditionsfrombeingsatisfied. pectation values of a problemsolver’s performanceon the metric across a problem set (MARKOVITCH~ SCOTT 4. Modeling CRL and CBR Systems 1993). Tocomputethe utility of a changeto the system’s To compare CBRand CRLsystems, we must develop knowledgebase with respect to somemetric, wewant to computational models of these systems and compare computethe costs that the systemwill incur for different learning andproblemsolving in each. problemsweightedby the probability that the systemwill The baseline for comparisonof the computational actually encounter those problems. Thus, utility is a model approach is the unguided problem solver. Unfunction not only of the learned item but also of the guided problemsolvers (UgPS)are a class of problem learning system, the problemset, problemdistribution, solvers that operate without search control knowledge. and the evaluation metric. The utility problemoccurs The algorithm of a UgPSis a knowledge-free weak whena learning system makesa change to its KBwith the goal of improvingproblemsolving utility on some Problem Space metric by a calculated improvementFc, but whichhas Initial State~ ~ the side effect of degradingproblemsolving utility for I another (possibly identical) metric by someactual Difficulty d= ’:~ii Solution Path amountFa that outweighsthe savings(i.e., Fc<Fa). path length 3.3. Dissecting the Utility Problem ,:,:,;, In general, utility problemsare not global, emergent Goal State properties of computationbut can instead be tied to Branching. Fact.or b= specific interactions betweenthe CA,the KB,and the average numcer orcnolcea pernode performancecharacteristics of the HA.In a CRLsystem, Figure 1. UnguidedSearch the interaction of interest is the relationship between 37 methodguaranteedto find a solution if one exists, such as breadth-first search; the only knowledge it uses is its operatorlibrary. For a given problemp with a solution of path length d operating in a search space with branchingfactor b, a UgPSwill examineba nodes during search (see Figure1). Thenumberof nodesthat the system expandsis termedthe complexity of a problemand is denotedCp. Becausethe UgPSsolves problemsin exponentialtime, it can serve as a "baseline"against which moreefficient learning systemscan be compared.Much of "intelligence" can be viewedas attemptsto reducethis combinatorialexplosionthroughthe use of heuristics or other techniques (NEWELL ~ SIMON1975; RAM& HUNTER 1992; SCHANK & ABELSON 1977). 4.1. Control-Rule Problem Solvers Learning systems improve over the UgPSby finding waysto reduce or eliminate search. A CRLsystem reducessearch by retrieving and applyingcontrol rules at eachstate it visits duringproblemsolving, givingit the ability to select or reject states. Thiscontrolknowledge is a completelydifferent kind of knowledgethan operator knowledge and must be stored in a separate control rule library. If a system’scontrol rule library is emptyand ot .C-. PROBLEM Elaborated Problem~~-n~::~ Problem Old Case & / ProUam & "-,,.I ~ed Case (so’ruuon) ~ problem Bolutl~ ~ & Outcome@ ~. Costof CaseRetrieval= Rc / ~ ~ I Old rSolution Pstfl lk,, Cost of ~1(,~t,.,c..) ~ ~ .=~..on I I /~N /~,:~ ~::::."2 ::: ~’~ ~ I New Solution Pa~ Cpffi TOtel of Nodes ~AdaptedCase F.Jmrnlned Inlq~h ,S~ce ~(SOfUtlon) Figure3. Searchin the Spaceof ProblemPaths solution path is achieved(see Figure 3). Oncethe new solutionis found,it is storedin the case library indexed for futureretrieval. 5. Analyzing Retrieval Costs Ouranalysis of these modelsfocuseson retrieval costs in CRLand CBRsystems---howmanyretrievals are made, and howmuchdoes each retrieval cost? Retrieval is ofProblem Space ten cited as the core source of powerfor CBRsystems, Initial State------~-¢~ ....,._ A PRlication of Control .Rules yet the retrieval cost is a critical factor in the swamping Keauces ~pace ~earonea utility problem.Ananalysis of retrieval costs in CRLand CBRsystemsbefore and after learning reveals interestSolution Pa~~,, . /~~--~Potential Search Space: Cp ing differencesin howeachdeals with retrieval. paceActually Searched:Cp’ For our analysis, wewill assumethat both the CRL Total Sav ngs of and CBRsystemsoperate on the sameproblemset, and _ . ~ .............. ~oal State--’--GuioedSearch: Cp- Cp’ that their problemspaces are defined by the sameoperator library. Wefurther assumethat the costs of adaptaFigure 2. SearchGuidedby Control Rules tion operatorsare roughlyequivalentto those of regular operators; this assumptionmayor maynot be true for control rules are not available, the problemsolverresorts particular systemsbut is reasonablefor this analysis. to blind search. Oncea solution path has been found, Wedefine a basic operation of retrieval, R, which correct decisionscan be cachedin the library as control rules that will guidethe systemin the future (see Figure chooses an item in a knowledgelibrary based on some 2). This model, while simplified, approximates many matchingfunction. In general, for a given hardwarearchitecture HA,the cost of retrieval for a knowledge liexisting systems,including Soarand Prodigy. braryi, denotedRi, is a functionof boththe library i and 4.2. Case-Based Reasoners the item to be retrieved, r: Ri =fir, i). For a serial hardware architecture, HAs,the most importantvariable in Case-basedreasoning is primarily experience-based; whena CBRsystem encounters a newproblem, it checks this cost function is the numberof items in the knowledge library, Ki. Wewill approximatethis serial cost its case library of past problemsolving episodes, or cases, lookingfor a similar casethat it can adaptto meet function with Ri = cKi, wherec is a constant multiplier that approximatesthe (nearly) linear cost function for the needs of the new problem. Our model of CBRhas matchingon serial systemslike 3Has. twoprimaryknowledge libraries: the case library itself, Theparticular interaction wewill examineis the reindexed so that the most appropriate case can be retrieved in newproblem-solvingsituations, and an adap- trieval time interaction: the relationship betweenthe retrieval operation Ri, knowledgelibrary size Ki, and tation library that stores adaptationoperatorsthat are runningtime t. Becausethe learning operations in CBR used to transform the cases once they are retrieved. and CRLsystemshavethe effect of increasing the size of Whena CBRsystemis presented with a problem, it retrieves an appropriate past case based on the problem’s knowledgelibraries in the system, their learning modules coupledwith the retrieval time interaction formpofeatures, its goals, andthe indicesit has in its case library. Then, the case is adaptedby performingsearch in 3 Anindexed or parallel memorysystemmight improveon this retrieval the space of problempaths: the adaptation operators transformentire paths into newpaths until a satisfactory function (or, then again, might not, depending on the domain; see ~ DOORElqI3OS 1993 and TAMBE ET AL, 1991), 38 tential root causes of the utility problem. Now,let’s examine the dynamics of learning and retrieval in CBR and CRLsystems and attempt to establish the threshoM conditions for the utility problemin each. 5.1. Retrieval in Control-Rule Problem Solvers In its initial state, without control rules, a CRLsystemis equivalent to the UgPS.It searches Cp states, retrieving a set of operators at each step with a cost of Ro. Nocontrol rules exist, so the cost of control-rule retrieval Rc = O. Thus, the total cost in retrievals of the initial systemis CpRo.After the systemhas learned a set of control rules, it has the capacity to guide its search. The number of states searched is reduced to Cp’, where Cp’ < Cp. However, in addition to retrieving a set of operators, it also needs to retrieve control rules at each step; thus, the cost for solving a problemrises to Cp’(Ro+Rc). The expected savings that guided problem solving brings are the costs of the states that the problemsolver avoids: (Cp - Cp)Ro. The added costs are the costs of matching the control rules at each step: Cp’(Rc’ - Rc) cp’ARc. (Note: because Rc = O, ARc = Re). The utility problem will arise whenthe added costs exceed the expected savings. Thus, the threshold condition is (Cp Cp)Ro < Cp’ARc; in other words, the swampingutility problem arises when the added cost of retrieval outweighs the benefits of individual rules. But will this threshold condition ever be met? In the limit, the maximumsearch reduction is to a single path (Cp’ = d), and operator retrieval costs axe constant (Ro’ = Ro) since the library of operators the system uses does not change in size. The maximumexpected savings possible for any problem are thus (Cp d)Ro. In contrast, the cost of retrieving control rules (Re) increases without boundas the control base Kc increases in size; in the limit, the added costs associated with a nflebase are dRc’ = dc(Kc) and thus can outweigh the maximumpossible savings. Therefore, the threshold conditions can be met and the CRLsystem will encounter the utility problem. These results indicate that swampingis a function of the potential speedupof learned items, the cost function of retrieval (which is itself dependent on retrieval strategies and machine architecture), and the numberof items a system needs to learn. If the system converges on a bounded set of learned items and the hardware slowdownnever approaches the utility of those items, the 4 If the learned items are system will never be swamped. of low utility, or if the learner never converges on a bounded set, as might be the case for an open-world or 4 For example, on a parallel machinewith a logarithmic cost ~nction Ri = c(log Ki), the threshold condition (Cp-Cp~Ro< Cp’c(log Kc) may never be met in a closed-world domainin which a small set of knowledge items learned by rote are adequatefor performance.If the learning system successfully converges on a small enoughset, the logarithmic slowdown will be negligible comparedto the potential savings. This condition can arise on serial architectures as well, but becausethe cost function is linear in the size of the knowledge base the constraints on the size of the learned set are muchmoresevere. 39 multidomain system, then the swamping problem can eliminate the benefits of the learned rules. 5.2. Retrieval in Case-Based Reasoners To analyze utility effects in CBRsystems, we need to measure the performance of a CBRsystem as it learns. To provide a basis for this measurement,we assume that a CBRsystem that does not have an appropriate case in memorycan resort to somemethod(e.g., adaptation of "null case" or "from-scratch" reasoning) that is costequivalent to the UgPS. Many existing CBRsystems have such a last-resort method (e.g., KOTON 1980, KOLODNER & SIMPSON1988). A CBRsystem that resorts to null-case adaptation beginning with no experiences must still incur the cost of retrieving the null case (Re) and then search the space of problempaths until the case has been adapted into a satisfactory solution. Under our earlier assumptions, the total numberof paths the system examinesis Cp, and one adaptation retrieval (Ra) occurs per step. Thus, the total cost of case adaptation before learning is Rc + CpRa. After the system has learned a library of cases, it will still need to retrieve a case but each case will require muchless adaptation, reducing the numberof paths examined to Cp’ where Cp’ << Cp. The cost of retrieving cases will increase to Re’ where Re’ >Re. Thus, the total costs are Re’ + ¯CpPRa TOevaluate these results we must again examine the benefits and costs of case retrieval. The expected savings are the costs of the states that the problemsolver avoids: (Cp - Cp) Ra, while the added costs are the increased costs of retrieval of cases Re’ - Re = ARc. In the limit, the cost of retrieval increases without boundas the case library increases in size: Re’ = c(Kc). However,as we approach the limit the case library contains manyappropriate cases and little adaptation needs to be dono-perhaps only one or two steps. In general, wheneverthe threshold condition (Cp - Cp) Ra < AReis met, the cost of retrieval outweighs the benefits of case adaptation; under these conditions, CBRsystems will be swamped. 5.3. Advantages of Case-Based Reasoning Whilethis analysis reveals that both control-rule learners and CBRsystems can suffer from the swampingutility problem, it also reveals that CBRsystems have important advantages over CRLsystems. The primary advantage CBRsystems have over control-rule problemsolvers is that cases are retrieved only once during the lifetime of problem solving. For a control rule problem solver to avoid swamping,the increase in cost of retrieval of a control rule mustbe less than the fraction of total states that the system avoids in guided problem solving times the cost of an operator:ARc < Ro(Cp- CpS/Cp’. For a CBRsystem, on the other hand, the increase in cost of a case retrieval must be less than the cost of the numberof adaptation steps avoided: ARc < Ra(CP- C )P ¯The missing C ’p term in the denominator of the CBRequation arises because the increased cost of retrieval of control rules are incurred at each step in the search space, whereasthe increased cost of case retrieval is incurred only once during problem solving for a CBR system. In other words, CBRsystems amortize the cost of case retrieval across all adaptations, making them muchmoreresistant to increases in’ retrieval costs than CRLsystems. This amortization also makes CBRmore amenable to solutions to the swampingproblem, such as deletion policies or indexing schemes. In order to be effective, any coping strategy needs to reduce retrieval time to the point that the threshold conditions are never satisfied. For CRLsystems, this upper limit is Re’ < Ro(Cp Cp)/Cp’; for CBRsystems, this upper limit is Rc’ < Ra(Cp - Cp%a muchhigher (and hence muchless stringent) limit on the maximum time retrieval can take for system to be guaranteed to avoid swamping. 5.4. Future Comparative Analysis The above analysis does not close the bookon the utility problem in CBRand CRLsystems. There are a number of other differences that can contribute to the utility problem, including differences in operator costs, degree of search reduction, and knowledgebase size. Normalproblem solving operators and adaptation operators do not necessarily have comparable costs. A traditional problem solving operator makesa change to a single state, while an adaptation operator can potentially make several changes to a case. Thus, the same reduction in Cp’ in CBRand CRLsystems could provide different degrees of savings because the individual steps being saved maynot cost the same. Another difference between CBRand CRLsystems is that they are not guaranteed to produce the same reduction in C/after being exposedto the sameset of training examples. While both types of systems are senstitive to the structure of the domain, search reduction in CRL systems depends on the learning and matching policy for control rules, but in CBRsystems depends on the indexing schemeand similarity metric used for case retrieval. Therefore, the same training set might produce different search reductions for CRLand CBRsystems. Furthermore, depending on the learning biases chosen, different systems could extract varying numbersof cases or rules from the same set of examples. Therefore, even if the search reduction is the same, there is no guarantee that the size of the systems’ knowledgelibraries will be comparable, and hence no guarantee that their performance on the utility problemwill be identical. Providing a principled account of the effect contributing factors like these have on the utility problemis the primary goal of our research. While we believe we have identified one factor -- amortization -- which contributes to the performance differences of CBRand CRL systems, it is only one contributing factor amongmany that determine the systems’ respective performances. Accountingfor other contributing factors, such as operator costs, degree of search reduction, and knowledgebase size, will allow a principled comparison of CBRand CRLsystems performance on the utility problem. For example, there have been few attempts to systematically evaluate the cost-utility tradeoffs in CBR 4O systems with very large case libraries. However,reports of the speedup provided by cases (e.g., KOTON 1989) and control rules (e.g., TAMBE ETAL1990) suggest that cases can provide large improvements, up to an order of magnitude greater than the speedups provided by control rules. It is possible that the utility of eases maybe high enough to allow CBRto avoid swamping, but it is not clear whetherthis will alwaysbe the case. 6. The Bottom Line CBR’spatterns of retrieval makeit resistant to the utility problem. Becausethe cost of case retrieval is amortized over many adaptation steps, ideal CBRsystems suffer less severely from the same overhead and are more amenable to coping strategies than CRL systems. This analysis suggests several future lines of comparativeresearch on CBRand CRL,such as operator costs, degree of search reduction and size of knowledgebases. References Doorenbos,R.B. (1993) Matching100,000 LearnedRules. Proceedingsof the EleventhNationalConferenceon Artificial Intelligence, p290-296.AAAI Press/MITPress. Etzioni, O. (1992) AnAsymptoticAnalysis of SpeedupLearning. In MachineLearning:Proceedingsof the 9th International Workshop,1992. Francis, A. and Ram,A. (1994). ComputationalModelsof the Utility Problemand their Applicationto a Utility Analysisof Case-BasedReasoning. Technical Report GIT-CC-94-24, Collegeof Computing, GeorgiaInstitute of Technology. Holder, L.B.; Porter, B.W.;Mooney,R.J. (1990). Thegeneral utility problemin machinelearning. In MachineLearning: Proceedingsof the SeventhInternationalConference,1990. Kolodner,J.L. &Simpson,R.L. (1988). TheMediator: A case study of a case-basedreasoner. GeorgiaInstitute of Technology, Schoolof Informationand ComputerScience, Technical Reportno. GIT-ICS-88/11. Atlanta, Georgia. Kolodner, J.L. (1993). Case-basedReasoning. MorganKaufmann,1993. Koton,P. A. (1989). A Methodfor Improvingthe Efficiency Model-BasedReasoningSystems. Laboratory for Computer Science, MIT,Cambridge,MA.HemispherePublishing. Markovitch,S. and Scott, P.D. (1993) InformationFiltering: Selection Methodsin Learning Systems.MachineLearning, 10: 113-151. Minton,S. (1988). Quantitativeresults concerningthe utility of explanation-basedlearning. In Proceedingsof the Seventh NationalConference on Artificial lntelligence, 1988. Minton,S. (1990). Quantitativeresults concerningthe utility of explanation-based learning. Artificial Intelligence, 42(23), March1990. Newell, A. &Simon, H.A. (1975). ComputerScience as Empirical Inquiry: Symbols and Search. Reprinted in Haugeland,J., (ed.). MindDesign:Philosophy,Psychology, Arnficial Intelligence, chapter1, pp 35-66.MITPress, 1981. Ram,A. &Hunter, L. (1992) The Use of Explicit Goals for Knowledge to GuideInferenceand Learning.AppliedIntelligence,2(1 ):47-73. Schank,R. &Abelson, R. (1977). Scripts, Plans, Goalsand Understanding.LEA,Hillsdale, NJ. Tambe,M.; NeweU,A.; Rosenbloom,P. S. (1990). The Problemof ExpensiveChunksand its Solutionby Restricting Expressiveness. MachineLearning,5:299-348,1990.