From: AAAI-87 Proceedings. Copyright ©1987, AAAI (www.aaai.org). All rights reserved. perationality for Ex Richard M. Keller ,Rutgers University Department of Computer New Brunswick, New Jersey 08903 Internet: Keller@Rutgers.Edu an initial target Abstrdct enced a surge of interest ing methods called et al., 19861. In ities among explanation-based knowledge-intensive to empirical analysis large concept numbers particular, although the initial given to an explanation-based ( i.e., it covers the correct tion is in non-operational the description to improve methods perform of a single meth- of similarin- an in-depth, training example - typically a positive instance. That analysis involves first explaining why the positive training instance is an example of the concept to be learned (the target concept), and then generalizing the explanation in a principled manner so it is valid for a larger class of instances than the original instance. Finally, a description of that larger class of instances is extracted from the generalized explanation structure. The description constitutes a generalization of the original instance. A seeming paradox of the explanation-based paradigm is that in order to produce its final description of the target concept, the learning system must possess an initial description of that same concept. In fact, without “knowing” an initial description of the target concept, it would be impossible for the system to explain why the given training instance is an example of the target concept. So if 482 Machine Learning & Knowledge Acquisition cannot having the task target system the initial descripare at be used effectively There a concept methods example to learn system, descrip- this means by the learner is a significant description that gap by transforming, analogy the Informally, differ- and being able system is to or operationalizing, the of an explanation-based initial description. As a concrete ston et al’s concept description generally is correct set of instances), form. task performance. ence between learn- of training with existing knowledge that is unusable or impracticable into a form that is usable [Keller, 1983, Dietterich, 19861. Pn [Mitchell learning analysis is wrong for necessary The way to untangle the paradox is to acknowledge that learning can involve knowledge transformation, as well as knowledge acquisition. Explanation-based methods do not acquire “new” knowledge, per se, but rather transform has experi- methods a simple syntactic and differences stances, in a class of analytic explanation-based contrast ods, which perform learning is a prerequisite1 then why is learning tion? What is there to be learned? These questions the very heart of the “explanation-based paradox”. narrow the field of machine description methods, in the first place ? What to use it; years, concept explanation-based Operationality is the key property that distinguishes the final description lecarned in an explanation-based system from the initial concept description input to the system. Yet most existing systems fail to define operationality with necessary precision. In particular, attempts to define operationality in terms of “efficient instance recognition” tacitly incorporate several unrealistic, simplifying assumptions about the learner’s performance task and the type of performance improvement desired. Over time, these assumptions are likely to be violated, and the learning system’s effectiveness will deteriorate. We survey how operationality is defined and assessed in several explanationbased systems, and then present a more comprehensive definition of operationality. We also describe an implemented system that incorporates our new definition and overcomes some of the limitations exhibited by current operationality assessment schemes. In recent Science for illustration, consider Win- which uses explanation-based a description for CUP [Winston et al., 19831. In this case, an initial CUP description is given to the system, expressed in functional terms: “a CUP is an open, stable, operational mation liftable because to enable vessel.” This it does not contain a vision system description the necessary (the performance is noninforsystem) to improve its performance in recognizing CUPS. In order for the description to be useful in improving performance, the learning system must transform the functional composed tion into a structural description, the vision system is designed to recognize: light object with a handle, pointing concavity.” a flat bottom, descrip- of primitives “a CUP is a and an upward- Although operationality is the key property that distinguishes the final description learned in an explanationbased system from the initial target concept description. most existing systems fail to define operationality with necessary precision. Current attempts to define operationality ‘The initial target concept description may be represented in the learning system explicitly (’ IX., in terms of declarative structures) or implicitly (i.e., within the system’s procedures) [Keller, 1987b]. tacitly incorporate several unrealistic, tions that may not hold throughout In particular, ationality many explanation-based as an independent, erty of concept dynamic, on the learner’s of performance understanding sophisticated properly in dynamic, In what cisely property improvement and analyze learning as a search through Suppose D1 in Figure scription provided is a of descriptions, task desired. and the A more is necessary systems that real-world environments. we define operationality follows, operprop- operationality performance of operationality explanation-based treat assump- binary-valued Actually, continuous-valued pendent The role of operationality with respect to explanationbased learning is clarified by viewing explanation-based systems static, descriptions. of learning. simplifying the course how operationality detype thorough to construct can function more is assessed pre- in sev- eral explanation-based systems. Then we describe how the MetaLEX system [Keller, 1987b, Keller, 1987a] overcomes some of the limitations exhibited by current operationality assessment schemes, by incorporating our new definition of operation&y. explanation a common introduces some terminology framework to serve as a basis for our discussion of operationality. We begin by distinguishing its description. drawn from sionally between by a concept description, Two of instances. the universe considered Figure synonymous 1 illustrates figure depicts if they these the space be described by a given concept space represents from instance space, space. denote over are concept. space. of the concepts regions. between into Note that there a point in con- description D1 and D2 are two synonymous space. As descriptions, So when.there exist different the same concept, operationality defines preferring one description over another. ways to describe the criterion for process his formueffectively The through the concept space (Dl) and the final descrip- concept (6). ity definition describing for Explanation-based a of operationality explanation-based most systems commonly cited in is the following: Caakrena%Opera%iondi%y efaa.: A concept description is operational if it can be used efficiently to recognize the figure de- op- is insufficient learning. the same Below we review both describing concept C, which covers instances I1,12, and 13. However D2 is considered operational, whereas D1 is not. denote the space is partitioned in concept tion (D2) can At left, We D2 “concept characterization involves no searching the initial description In the center, descriptions The description This explanation-based concepts that into the space, equates a concept with its description, thereby masking the issue of operationality. Thus he effectively characterizes learning as a search through a concept space, not a concept and learning program. Each point in a unique set of instances drawn shown at right. cept space and the points the same relationships. operational and non-operational is a one-to-many correspondence illustrated, which is a predicate concept descriptions of all possible picts the space of all possible in concept a concept A concept represents a subset of instances A concept is denoted intensome universe. Dl we can criterion. between [Mitchell, 19821. A crucial distinction lation of the problem and ours is that Mitchell learning because aud establishes Then D2 as a solution termination of transforming de- and D2 erationalization” [Keller, 1987b, Keller, 19831. Mitchell first described learning as a search deac&ption section learned. for traversing as the search space. system, node in a search, as the means and operationality describing This description as the starting call the process description non-operational to an explanation-based is the final, operational node, the concept 1 is the initial, instances several systems that that operationality on our retrospective use explanation-based methods. defined so the following analysis et it denotes. is instantiated has not been eral of these systems, ins&cm of the concept how this definition de95 explicitly definitions in Note in sevare based and reconstruction. nc%icm-S%ruc%use s system, system the target concept is the set of drinking CUPS and the initial description is a functional description: “a CUP is any open, stable, liftable vessel.” Instances, tions of CUBS s tructud properties, etc. Therefore, in this system terms, from however, consist the real world, of physical expressed such as “flat” ) “handle”, an operational as a description descrip- in terms of “concave”, CUP description is defined stated solely in structural so it can be easily matched against instances. 2 [Mitchell et ad., 19861: In LEXB, the target concept is the set of USEFUL problem solving moves to apply during search. The initial description of USEFUL given to LEX2 states that “USEFUL diately or eventually to a solution.” moves lead immeInstances consist of calculus problem solving moves generated while solving tual. problems. TQ facilitate matching against instances, operational description Figure 1: csncepe acan description of USEFUL is defined in LEXB as a expressed in terms of the calculus features used to describe instances (e.g., “sin”, “3”, ‘“product”, etc.), or in terms of features easily derivable from them (e.g., “trigfunction”, “integer99 9 6‘po1ynomia199 9 etc.) S Keller 483 . o PRODIGY a variety including [Minton, of PRODIGY’s problem shop scheduling domain, formed 19861: This system the “efficient learns of target concepts related to problem solving, the USEFUL concept learned by LEXB. One into CLAMP, finished solving goods using etc. PRODIGY, under which applying SUCCESSFUL. Instances the machine-shop operators correspond environment description stance can learn to different is UNstates of the operators are recognition description, in the environment, tural, rect observability requirement “idle”, assures “busy”. that The di- UNSUCCESS- proper e GENESIS [Mooney and DeJong, 19851: In one of GENESIS’s application domains, the target concept corresponds to WEALTH-ACQUISITION-SCENARIO, and including a larger An operational number of of performance napping, from in the three systems stances, by the described GENESIS is not present in the in- but rather in terms of abstract schemata possessed system. A high-level description of WEALTH- ACQUISITION-SCENARIO recognition” because in GENESIS parses top-down facilitates the story stories “efficient understanding (i.e., instances) instance component efficiently in fashion. can be description. definition time is the oper- The including cost, way to eliminate the definition there are arbitrarily might be relevant to elegance, these simplicity, restrictive of operationality etc. assumptions is to redefine that system’s inition performance. Table of operationality. operationality There in the revised 1 gives our revised def- are two requirements on definition: usability and utd- dty. The usability requirement ensures that the description can be used by the performance system. This means that the description must be expressed corresponds to the original duced in [Mostow, 19811.) notion of operationality The utility requirement usability the description completes purposes, we can consider a specific subgoal problem attempts “similar” For our (along with the current processing state) as a training instance, the class of “similar” subgoals as SOAR’s target concept, and the generalized production’s conditions as its operational description. An operational description of the target con- cept is one that can be used to efficiently recognize whether a “similar” subgoal is true in a given processing state. In other words, in SOAR an operational production consists solely of conditions that can be easily evaluated without present in conditions initially problem solving, including that are evaluSOAR’s working memory, and conditions ated using chunks acquired during problem solving. The notion ployed in these of “efficient instance systems is a suitable defining operationality, as a general-purpose 484 recognition” emstarting point for but is in several ways inadequate definition. A basic problem is that Machine learning & Knowledge Acquisition or computable by the system. one step Table and in terms of capabili- chunking Each time SOAR by the system, in terms ties possessed mechanism. it in terms of the performance system that uses the learned concept description, and in terms of the criteria for evaluating e SOAR [Rosenbloom and Laird, 19861: In SOAR, a form of explanation-based learning is an integral part of its solving activity for a specific subgoal, the system to construct a generalized production to achieve subgoals without resorting to problem solving. for than struc- evaluating performance, including space efficiency. A description that is operational with respect to time efficiency may not be operational with respect to space efficiency. performance, Unlike description rather to use in evaluating of natural language text describing stories involving acquisition of wealth, such as stories involving inheritance, kidetc. be ationality. However, there are other types of ‘6efficiency9’ that may be just as appropriate or more appropriate for Furthermore, aside from efficiency, large numbers of other criteria that arson, gen- we might of novel cup designs functional SCENARIO consists of any sequence of actions that culminate in an agent’s acquisition of wealth.” Instances consist above, an operational description for stated in terms of the low-level features instance domain, is functional, from the abstract measure in- use of a concept Second, the “efficient instance recognition” used by most systems assumes that execution FUL conditions can be recognized quickly and operator application can be avoided in a real-time environment. an initial (although not explicitly stated) description the target concept is that “a WEALTH-ACQUISITION- of cups). of generation because generated uses, descrip- Although in recognizing cups (e.g., if we want to or in generating cups (e.g., if we are de- signing new types als and the machine-shop performance. instances”. in the CUP incor- how the final that the concept one typical are other interested either drink a beverage) the purposes equipment represents there must be phrased in of the raw materi- “temperature”, assumes For example, e&ion. tacitly about will be used to improve the definition applied. An operational description terms of directly observable features such as “shape”, definition assumptions tion will be used to “recognize like LATHE, operators in which concept First, are trans- for example, these recognition” several restrictive is a machine- in which raw materials POLISH, conditions domains instance porates further: (The of data usability known requirement must introtakes not only Operationallity Definition 1: Revisedl Given: QBA concept description CBA performance ayatem that makes use of the descrip- tion to improve l Performance of system performance objectives specifying improvement the type and extent desired operational is considered Tiren: the concept description if it satisfies the following - two requirements: 1. usability: formance 2. utility: mance the description must be usable by the per- system when the description system, the system’s prove in accordance is used by the perforperformance must im- with the specified objectives. be usable, description system, but also worth using. In particular, using the must improve the behavior of the performance as defined by its performance As an example objectives. of how this revised or dynamic. operationality defi- nition might be instantiated for existing explanation-based systems, consider once again Winston et al’s CUP domain. In this domain, the performance system might consist of a mobile robot searching for cups in a room. The performance objectives for the robot might be to improve the speed with which a CUP it can recognize description correspond able to easily the properties a structural for instance, an operational With For it must that property would the robot must such a description progresses, may become operational, due to changes in the performance curate assessment assessment system is made. and white camera. tions which specify a dynamic do not. namic because assessment suppose the notion situation. instead systems of operationalContinuing we are learning acquire descriptions with operational objective al’s system and in PRODIGY remains the course of learning, so these systems the number the system can generate. Now the revised definition correctly pinpoints an operational of cup designs operationality description as one expressed in terms of the functional design primitives known by the system, instead of structural primitives. subsystem [Utgoff9 generalization ever, that the set of operational cally adjust to changes As these b the set of is enlarged. in LEX2 when the be to increase of the or chunks, cups in a design context. In this case, the performance system might consist of a design system containing a library of functional design primitives. The performance might STABB is dy- in terms respectively. schemata descriptions term to the system’s section whereas and SOAR is defined or chunks, considered the set of operational about system. in the previous in GENESIS additional camera of operationality, operation&y ezistdng set of schemata definition, for the updated surveyed Assessment if the these same descriptions operational Some of the systems others of with a black However, with a color camera, as part a performance equipped For this system, any object descripcolor attributes should be considered should be considered perform An ac- on when the consider robot for recognition. non-operational were replaced is initially and vice versa, depends For example, of a mobile that environment. of operationality consisting as “specific- not be permitted to fit the learning the above example, As learning non-operational be used in the descrip- description. the revised ity adjusts by the robot, properties as well as “usable”, evaluate Therefore, gravity”, of object cups. can be de. sensory systems. Those properties to structural properties. For a CUP descrip- tion to be “utile”, tion. and retrieve to be “usable” be expressed in terms tected by the robot’s A dimension characterizing whether eVn&bility: operationahty assessment varies with time. Values: datdc Similarly, is augmented 19861 adds language. descriptions a new Note, how- in Winston et static throughout cannot automati- in the performance environment. e Granularity: A dimension characterizing the assessment measure produced. Values: binary or continuous. Most of the sessment systems surveyed of operationality: operational”. Rowever, either produce a binary “operational99 continuous-valued as- or “non- assessment has In this section, we discuss how operationality is assessed in explanation-based systems. Thus we draw a distinction distinct advantages over binary assessment because it allows the learning system to assess degrees of operational- between how operationality is defined and how it is evaluated in practice. Conceptually, each explanation-based ity. system contains which evaluates an operationality a concept sure of its operationality three dimensions which characterize and produces as output. - variability, Below, granularity, the operationality duced by an assessment procedure. ous systems’ assessment procedures is given in Table 2. (The table scribed in the previous section, system, aa.9essment procedure, description which is described a mea- we describe and certainty measurements A comparison - pro- includes the systems deas well as the MetaLEX in the next section.) where there exist several synonymous, (i.e., most effective) description. Additionally, continuousvalued assessment facilitates attempts to guide the search through concept of progress of vari- along these dimensions In situations operational descriptions of the target concept, a metric on operationality enables the system to learn the “best” PRODIGY valued description through is one assessment. sess how efficient space by providing a measure the space. system that In other a given features words, description continuous- PRODIGY can as- is for the purposes of recognition. The system bases a priori estimate of the matching this assessment on an costs associated with each machine;shop “observable” feature in the ment. Given two synonymous, the target concept, PRODIGY erational using the matching Certrainty: in the measurement ment procedure. cost estimates. A dimension produced Values: None of the systems environ- operational descriptions of evaluates which is more opcharacterizing confidence by the operationality unguaranteed surveyed can guarantee racjr of their assessment measurements, systems directly assess operationality. assess- or gumznleed. the accu- because none of the In other words, the Keller 485 systems do not actually test whether a given description is “efficient to use for recognizing instances’9. Instead, the systems base their assessments being assessed is contained of operational descriptions, on whether lem solving the description language within a pre-defined which is specified by the learn- The problem sumptions mance is that upon which invalid. system’s the the designer based capabilities change discussed an example from in the previous guage consists section, of descriptions solving as more chunks behavior SOAR’s are learned correctly deteriorate and matching costs lancondi- and thus However, are preSOAR’s over time increase. (This type of difficulty has been documented - with plans, rather than chunks - for STRIPS-type problem solvers involving any ar[Minton, 19851.) E ventually, descriptions bitrary chunked condition will no longer be efficient to use. At this point, it may be necessary to redefine operationality so that only the “most efficient” chunks are permitted within operational descriptions. But SOAR lacks the proper perspective over its own problem behavior to identify that the operational tion has become not automatically so changing solving/learning language defini- invalid over time. Moreover, SOAR can“recompile” a’new language definition, the definition requires human MetaLEX climbing algorithm from Table intervention. 2, the MetaLEX program is in a different equivalence class with respect to these three dimensions. The next section discusses MetaLEX and its treatment of operationality. efficiency The MetaLEX program [Keller, 1987b, successor to LEX2, designed to explore Keller, 1987a] is a the concept oper- ationalization paradigm introduced in [Keller, 19831. MetaLEX approaches essentially the same learning task as LEXP, but solves it using a different technique. In particMetaLEX set of USEFUL forward 486 problem search. operational and LEX2 target Both solving systems concept from the out operators used jectives performance and the other require SOLVER’s that problem formance learn a description moves start description to execute with the (“a USEFUL Machine learning & Knowledge Acquisition of the during same initial to search the on how MetaLEX description. guides efficiency establish concept, SOLVER benchmark calculus nonprob- per- over the description search. inserts that SOLVER’s problems. ob- improve while also maintain- for a description MetaLEX and observes The and effectiveness a metric the hill-climbing one involving description efficiency To assess operationality FUL system, effectiveness. an operational The measures space that involving solving ing its effectiveness. of the USE- description behavior During into on a set of SOLVER’s execu- tion, the USEFUL description guides expansion of problem solving moves: any move recognized as USEFUL is expanded, while other moves are pruned. SOLVER’s efficiency and effectiveness are monitored during the problem solving session. If SOLVER’s performance meets the Table 3: MetaLEX Operationality Definition Given: problem solving Performance problem Description moves of class of USEFUL to expand during SOLVER, system: solving Performance ular, both searches of traversing search a forward those problems Then: the USEFUL if it satisfies 1. usability: SOLVER 2. utility: search system objectives: Improve SOLVER’s efficiency on a set of benchmark calculus by X%, while maintaining its effectiveness MetaL V. descrip- of an operational description We do not describe the hill- or the Concept description: As is evident concept space in this paper. Instead, we focus assesses the operationality of a concept ified for the SOLVER As operational production target for ac- The definition of operationality used by MetaLEX is given in Table 3. Note that there are two objectives spec- result for clarification. involving will likely as- language no longer SOAR In contrast, to as- The state”), descrip- methods is used as a means description in the direction using a form of hill-climbing. as the perfor- over time. tions that have already been chunked, sumed efficient to use in recognition. problem the In fact 9 this is inevitable differ is in their region. The explanation the search space. pos- performance is that the original language definition identifies operational descriptions. Consider language original the systems Where to a solution an operational space, with the initial target concept description anchoring the search in the non-operational region and a training instance description anchoring the search in the operational descripfeatures the lan- of schemata with using a “compiled” sess operationality, may become composed tion. it into tion using explanation-based methods. In a sense, LEX2 conducts a bidirectional search of the concept description recognition. In a sense, the language is a “compiled” form of the designer’s knowledge about the performance system guage includes any description sessed by the system. eventually to transform complishing the transformation. LEX2 transforms the initial ing system’s human designer. This language includes only terms that the designer decides can be evaluated efficiently by the performance system in order to accomplish instance [Keller, 1987b]. For LEX2, that language includes tions expressed in terms of a variety of “syntactic” of calculus problem solving states. For GENESIS, move leads and attempt run-time problems in solving correctly description the following is considered operational two sequirements: the description is usable (i.e., evaluable) by and using the description, achieves an X% improvement in effectiveness. SOLVER’s without efficiency a deterioration established performance tion is considered fails to attain objectives, operational. the desired levels descrip- this research was provided perform=ce and DARPA contract the USEFUL If SOLVER’s of efficiency In the terminology of the previous section, MetaLEX’s operationality assessment method can be characterized as: - because operationality pends on the current state of SOLVER performance - because degree of effectiveness lems solved); assessment its performance However, MetaLEX’s expensive because each aLEX time a measure seconds) and the of benchmark assessment SOLVER, meets prob- assessment costs is required. by estimating system 1987a] shop, University perfortest. uary 1987. As such, operationality for an explanation-based system (or more generally, for a knowledge tion system) to “learn.” Yet current methods operationality tacitly depend on simplifying is associated with violated assumptions by defining operationality directly in terms of the performance system. The of operationality in MetaLEX may provide in- sight into how to construct general purpose explanationbased learning systems systems that are more sophisticated in their operationality assessment capabilities, and that function properly over time, and for tasks other than “efficient instance Technical recognition. [Minton, directing this encouragement this paper. research. Discussions ify the presentation provided Smadar and helpful with of PRODIGY formatting Kedar-Cabelli comments Steve and SOAR. assistance. Funding June Work- 1987. #ML-TR-7. CA, August Improving In learning. explanation-based Corvallis, 19821 September T. M. Mitchell. 1986. Generalization as search. 18(2):203-226, March Intelligence, et aI., 19861 T. M. Mitchell, S. T. Kedar-Cabelli. tion: a unifying view. of Proceedings of the Compilation, Oregon State Workshop on Knowledge University, 1985. the effectiveness 1982. R. M. Keller, and Explanation-based generaliza&earning, l(l), 1986. Machine [Mooney and DeJong, 19851 R. Mooney and G. DeJong. Learning schemata for natural language processing. CA, August [Mostow, pages IJCdI-8, 681-687, 19811 D. J. Mostow. of Task Heuristics thesis, Mechanical Transformation into Operational Procedures. PhD Comp.Sci.Dept., [Rosenbloom and Laird, @MU, 19861 1981. P. S. Rosenbloom Mapping explanation-based Laird. onto soar. In Proceedings AddI-86, phia, [Utgoff, Los Angeles, 1985. PA, August [Winston Philadel- Machine Learning ofInductive Academic, et al., 19831 AAAI, 1986. 19861 P. E. Utgoff. Bs’as. Kluwer and J. E. generalization Hingham, P. H. Winston, and M. Lowry. Learning definitions, In Proceeds’ngs AdAI-89, D.C., August 1983. MA, T. 0. physical examples, 1986. Binford, B. descriptions and precedents. pages 433-439, Washington, provided on earlier Minton in context. Selectively generalizing plans In Proceedings IJCd I-9, Los Angeles, from functional and to Jack guidance in 1983. Learning Irvine, 19861 S. Minton. Katz, Thanks go to my thesis advisor, Tom Mitchell, Mostow, both of whom provided invaluable August learning &facAine Report problem-solving. In Proceedings transformafor assessing performance assumptions that likely will be violated as learning progresses. The MetaLEX program circumvents the problems handling Concept of Galifornia, 19851 S. Minton. [Mitchell system. it means D.C., conAAAI- 1987bj R. M. Keller. The Role of Explicit Contextual Knowledge in Learning Concepts to Improve PhD thesis, Rutgers University, JanPerformance. Artificial Operationality is the key feature that distinguishes the output concept description from the input concept description by re-expressing In Proceedings [Keller, [Mitchell, in an explanation-based at the heart of what R. M. Keller. pages 596-599, Met- a system recognition. 89,pages 182-186,Washington, for must be tested in lieu of executing #DCS83-51523 Learning at the knowl1(3):287-316, 1986. 19831 R. M. K e 11er. Learning cepts for efficient [Minton, is also extremely system assessment possible by whether objectives. method the performance these mance whenever is accomplished and observing the stated operationality reduces yields (in CPU (in number - because executing [Keller, In Proc. 4th International and e “guaranteed [Dietterich, 1986] T. G. Dietterich. edge level. Machine Learning, [Keller, of the degree of efficiency directly assessment deand the current objectives; @ “continuous” grant or effective- ness, MetaLEX can evaluate how far from operational the description is, and can assess whether the operationalization search is headed in the right direction. ~9 “dynamic” by NSF #N00014-85-K-0116. drafts helped of clar- Chun Liew to support Keller 487