AI Magazine Volume 22 Number 2 (2001) (© AAAI) An ...

■ This article presents a learning agent shell andmethodology for building knowledge bases andagents and their innovative application to thedevelopment of a critiquing agent for militarycourses of action, a challenge problem set by theDefense Advanced Research Projects Agency’sHigh-Performance Knowledge Bases Program. Thelearning agent shell includes a general problem-solving engine and a general learning engine for ageneric knowledge base structured into two maincomponents: (1) an ontology that defines the con-cepts from an application domain and (2) a set oftask-reduction rules expressed with these con-cepts. The development of the critiquing agentwas done by importing ontological knowledgefrom CYC and teaching the agent how an expertperforms the critiquing task. The learning agentshell, the methodology, and the developed criti-quer were evaluated in several intensive studies,demonstrating good results.

Agreat challenge for AI is the developmentof theories, methods, and tools thatwould allow users that do not have

knowledge engineering or computer science

experience to build knowledge bases andagents by themselves. We believe success in thisarea will have an even greater impact on oursociety than the development of personal com-puters. Indeed, if personal computers allowedevery person to become a computer user, with-out the need for special training in computerscience, solutions to this AI challenge wouldallow any such person to become an agentdeveloper. Agent development by typical com-puter users would lead to a large scale use ofcomputers as personalized intelligent assis-tants, helping their users in a wide range oftasks. The key issue is that the development ofsuch an agent should be as easy for the user asit currently is to use a word processor.

In this article, we present recent progressmade in the George Mason University Learn-ing Agents Laboratory toward this goal. First,we introduce the concept of a learning agentshell as a tool to be used directly by a subject-matter expert (SME) to develop an agent.Then, we present the DISCIPLE family of learn-ing agent shells and the successful application

Articles

SUMMER 2001 43

An Innovative Application from theDARPA Knowledge

Bases ProgramsRapid Development of a

Course of Action Critiquer

Gheorghe Tecuci, Mihai Boicu, Michael Bowman, and Dorin Marcu

Commentary by Murray Burke

Copyright © 2001, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2001 / $2.00

AI Magazine Volume 22 Number 2 (2001) (© AAAI)

clude the article with a discussion of thisresearch.

Learning Agent ShellWe have introduced the concept of a learningagent shell as a new class of tools for rapiddevelopment of practical end-to-end knowl-edge-based agents by domain experts with lim-ited assistance from knowledge engineers(Tecuci et al. 1999). A learning agent shellincludes a general problem-solving engine and

of one member of this family, DISCIPLE-COA, tothe Defense Advanced Research ProjectAgency’s (DARPA) High-Performance Knowl-edge Bases (HPKB) Program’s course-of-action(COA) challenge problem. We describe thechallenge problem and the DISCIPLE-COA shelland methodology used to build the COA cri-tiquing agent. We then present the results ofDARPA’s evaluation of the developed toolsand COA critiquers. We also briefly presentthe results of a separate knowledge-acquisi-tion experiment with DISCIPLE-COA. We con-

In his invited talk at the 1993 NationalConference on Artificial Intelligence,Edward Feigenbaum compared the tech-nology of a knowledge-based computersystem with a tiger in a cage. Rarely doesa technology arise that offers such awide range of important benefits. How-ever, this technology is still far fromachieving its potential. This tiger is in acage, and to free it, the AI research com-munity must understand and removethe bars of the cage.

We now know that a knowledge-based system needs a great deal ofknowledge to be truly useful. However,building a large and high-performanceknowledge base is still a long, ineffi-cient, and error-prone process. Respond-ing to this problem, the DefenseAdvanced Research Projects Agency(DARPA) has sponsored a sequence oftwo programs for the development ofthe second generation of knowledge-based systems science and technology:(1) the High-Performance KnowledgeBases Program (HPKB FY97-99) and (2)the Rapid Knowledge Formation Pro-gram (RKF FY00-03).1,2 The goal of theHPKB Program was to produce the tech-nology needed to enable system devel-opers to rapidly construct large knowl-edge bases that provide comprehensivecoverage of topics of interest, arereusable by multiple applications withdiverse problem-solving strategies, andare maintainable in rapidly changingenvironments. The tasks of knowledgerepresentation and acquisition are toodifficult to be started from scratch eachtime a new knowledge base needs to bebuilt. Therefore, this program supportedthe development of methods for creat-ing knowledge bases by selecting, com-posing, extending, specializing, andmodifying components from a library ofreusable ontologies, common domaintheories, and generic problem-solving

strategies. In addition, it supported thedevelopment of methods for rapidlyextracting knowledge from natural lan-guage texts and the World Wide Weband for knowledge acquisition from sub-ject matter experts (SMEs). An impor-tant emphasis of the HPKB Program wasthe use of challenge problems, which arecomplex, innovative military applica-tions of AI that are intended to focus theresearch and development efforts andmeasure the effectiveness of alternativetechnical approaches. The participantscollaborated in the development ofknowledge-based systems for solvingthese challenge problems. These systemswere the subject of intensive annualevaluations during which the complete-ness and correctness of the developedknowledge bases were measured as wellas the time required to build the knowl-edge bases and the ease of modifyingthem to assimilate new or changedknowledge. The challenge problems forthe first part of the HPKB Program andthe corresponding evaluation resultswere presented in the Winter 1998 issueof AI Magazine (Cohen et al. 1998). Thesecond part of the HPKB Program wasbased on even more complex challengeproblems. One challenge problem wasan extension of the year-one crisis-man-agement problem, requiring rapid devel-opment of a large knowledge base con-taining broad but relatively shallowknowledge (such as general knowledgeof countries, politics, geopoliticalevents, economics) necessary to discov-er and understand information aboutnascent and emerging crises from a widerange of potential information sources.Two teams, one composed of Teknowl-edge, Cycorp, Textwise, and Northwest-ern University, and the other composedof Science Applications InternationalCorporation (SAIC), SRI International,the Massachusetts Institute of Technolo-

gy (MIT), Stanford University, andNorthwestern University, developed twoend-to-end integrated systems that wereevaluated by Information Extractionand Transport Inc. (IET), the challengeproblem developer, in the summer of1999. Both systems demonstrated highperformance through knowledge reuseand semantic integration and created asignificant amount of reusable knowl-edge. The other challenge problemrequired the rapid development ofknowledge bases containing compre-hensive battlefield knowledge (forexample, terrain characteristics, forcestructures, military organizations, troopmovements, military strategy). Theproblem was to assess various aspects ofmilitary courses of action (COAs), suchas their viability, their correctness, andtheir strengths and weaknesses withrespect to the principles of war and thetenets of army operations; to justify theassessments made; and to proposeimprovements. This challenge problemwas solved by developing a complexend-to-end system integrating comple-mentary technologies developed by dif-ferent research groups. Teknowledge,Northwestern University, and the Uni-versity of Edinburgh’s Artificial Intelli-gence Applications Institute (AIAI)developed the translation and fusionmodule that interpreted and combinedthe information included in the COAsketch and the COA statement and gen-erated an internal representation of theinput COA based on a CYC ontology.Four critiquers, each developed by a dif-ferent team (a joint team from Teknowl-edge and Cycorp, the DISCIPLE team fromGeorge Mason University, the EXPECT

team from University of Southern Cali-fornia Information Sciences Institute[USC/ISI], and the LOOM team fromUSC/ISI shared the generated represen-tation of the COA and critiqued differ-

Articles

44 AI MAGAZINE

a general learning engine for building a knowl-edge base consisting of an ontology and a set ofproblem-solving rules. The process of develop-ing a knowledge-based agent for a specificapplication relies on importing ontologicalknowledge from existing knowledge reposito-ries and teaching the agent how to performvarious tasks in a way that resembles how anexpert would teach a human apprentice whensolving problems in cooperation (figure 1).This process is based on mixed-initiative rea-soning that integrates the complementary

knowledge and reasoning styles of the SME andthe agent and on a division of responsibility forthose elements of knowledge engineering forwhich they have the most aptitude, such thattogether they form a complete team for knowl-edge base development.

The concept of a learning agent shell is anextension of the concept of expert system shell(Clancey 1984). As an expert system shell, itincludes a general inference engine that can bereused for multiple applications. In addition, alearning agent shell exhibits an organization of

ent aspects of it. The answers from eachcritiquer were displayed by a solutionviewer developed by SAIC. The integrat-ed system and its individual compo-nents were evaluated by Alphatech asexhibiting performance at the level of asubject-matter expert.

The HPKB Program emphasized thedevelopment of very large knowledgebases by knowledge engineers, demon-strating the utility of large knowledgebases and the feasibility of large-scalereuse. It has produced reusable knowl-edge libraries, including an upper ontol-ogy and middle theories for crisis andbattlefield reasoning. The follow-on RKFProgram emphasizes the developmentof knowledge bases directly by thedomain experts. Its central objective isto enable distributed teams of SMEs toenter and modify knowledge directlyand easily, without the need for priorknowledge engineering experience. Theemphasis is on content and the meansof rapidly acquiring this content fromindividuals who possess it, with the goalof gaining a scientific understanding ofhow ordinary people can work with for-mal representations of knowledge.Therefore, the program’s primary re-quirement is the development of func-tions that enable SMEs to understandthe contents of a knowledge base, enternew theories, augment and edit existingknowledge, test the adequacy of theknowledge base under development,receive explanations of theories con-tained in the knowledge base, anddetect and repair errors in content.Because of the complexity of these tasks,the approaches developed in the RKFProgram exploit the synergies amongcomplementary AI technologies, such asnatural language discourse processing,problem solving and learning by analo-gy, and commonsense reasoning.

RKF is organized in a manner similar

to HPKB, with challenge problemsadministered by an evaluation contrac-tor (IET assisted by Veridian Pacific-Sier-ra Research (PSR) and George MasonUniversity Institute for Biosciences,Bioinformatics, and Biotechnology) as abasis for formal evaluation of the tech-nology provided by the developers.There are two integrated teams and sev-eral component technology developersthat contribute to one or both of them.One team is coordinated by Cycorp andincludes research groups from USC/ISI,University of Edinburgh, Teknowledge,SAIC, and Northwestern University. Theother team is coordinated by SRI andincludes research groups from the Uni-versity of Texas at Austin, USC/ISI, Stan-ford University, Boeing, the Universityof Massachusetts at Amherst, Veridian/PSR, and Northwestern University. Thecomponent technology developersinclude Stanford University, MIT, North-western University, George Mason Uni-versity, University of West Florida, andPragati.

The RKF challenge problems aredesigned to test SMEs’ abilities to useRKF tools to build knowledge bases intwo main categories related to under-standing biological weapons: (1) text-book knowledge and (2) expert knowl-edge. The textbook knowledge challengeproblem covers standard textbooks inundergraduate biology and is designedto drive the development of the technol-ogy that will enable an SME to developknowledge bases for domains in whichknowledge is already relatively well orga-nized, self contained, or comprehensiveand for which there might exist recog-nized tests of subject understandingdefined independently of the RKF Pro-gram. These tests will be based on ques-tions appearing in the textbooks them-selves or in associated academic tests forthe relevant subject. The expert knowl-

edge challenge problem, however, coverspractical, hands-on tasks for militaryanalysis regarding defensive biologicalwarfare. This problem is designed to dri-ve the development of the technologythat will enable the professional SME todevelop knowledge that is task oriented,is not necessarily written down in books,and is germane for an application ofnational security importance. No appro-priate, independently defined test existsnow, so it will be necessary to defineappropriate tests in the context of theRKF Program.

This article presents the DISCIPLE tech-nology developed by the George MasonUniversity Learning Agents Laboratoryand its application to the COA chal-lenge problem. The developed tool,called DISCIPLE-COA, is an importantresult of the HPKB Program that is repre-sentative of the many accomplishmentsof the program. DISCIPLE is now evolvingfrom a tool for knowledge engineers to atool that can be used directly by SMEs,as demonstrated by its successful knowl-edge-acquisition experiment performedat the U.S. Army Battle Command BattleLab. In this sense, DISCIPLE is a good illus-tration of the transition from HPKB toRKF.

—Murray Burke

Notes1. Murray Burke, Rapid Knowledge For-mation (RKF) Program Description,dtsn.darpa.mil/iso/ programtemp.asp?mode=3.

2. David Gunning and Murray Burke,High-Performance Knowledge Bases(HPKB) Program Description, dtsn.darpa.mil/iso/index2.asp?mode=9.

Articles

SUMMER 2001 45

cation in a given domain, but they are evenspecific to a particular SME. Consider, forexample, an agent that assists a military com-mander in critiquing courses of action (COAs)with respect to the principles of war and thetenets of army operations (an agent that will bedescribed in more detail in this article). Therules will identify strengths and weaknesses ina military COA and will obviously be domainspecific. Moreover, they are likely to includesubjective elements that are based on the expe-rience of a specific military expert. Definingsuch problem-solving rules is a complexknowledge engineering task. Therefore, thelearning agent shell should include mixed-ini-tiative methods for learning such rules from anatural interaction with an SME.

The DISCIPLE ShellOver the years, we have developed a series ofincreasingly more advanced learning agentshells from the DISCIPLE family. The problem-solving engine of a DISCIPLE agent is based onthe general task-reduction paradigm. In this

the knowledge base into an ontology that spec-ifies the terms from a particular domain and aset of problem-solving rules expressed withthese terms.

The ontology is the more general componentof the knowledge base, characteristic to anentire domain, such as medicine or the mili-tary. A domain ontology specifies terms that areuseful in a wide range of different applicationsin a domain. For example, a military ontologywould include specifications of military unitsand military equipment that are likely to beincluded in the knowledge base of any agentdeveloped for a particular military application.Moreover, there is wide agreement in anymature domain on the basic terms of thisdomain, allowing one to reuse ontologicalknowledge that was previously developed tobuild a new knowledge base. As a consequence,a learning agent shell should include modulesfor importing ontological knowledge fromexisting knowledge bases.

The problem-solving rules represent the spe-cific component of the knowledge base. Therules are not only specific to a particular appli-

Articles

46 AI MAGAZINE

Figure 1. Interacting with a Learning Agent Shell.

The expert teachesthe agent to perform various tasks in a way that resembles

how the expert would teach a person.

ProblemSolving

The agent learnsfrom the expert,

building, verifyingand improving itsknowledge base

Learning

Ontology+ Rules

paradigm, a task to be accomplished by theagent is successively reduced to simpler tasksuntil the initial task is reduced to a set of ele-mentary tasks that can be performed immedi-ately. The ontology of a DISCIPLE agent is basedon the frame knowledge model of the openknowledge base connectivity (OKBC) protocol.OKBC has been developed as a standard foraccessing knowledge bases stored in differentframe-representation systems (Chaudhri et al.1998). It provides a set of operations for ageneric interface to such systems. There is alsoan ongoing effort to develop OKBC servers forvarious systems, such as CYC (Lenat 1995),ONTOLINGUA (Fikes, Farquhar, and Rice 1997),and LOOM (MacGregor 1991). These servers arebecoming repositories of reusable ontologiesand domain theories and can be accessed usingthe OKBC protocol. The use of the OKBCknowledge model for the DISCIPLE ontologyfacilitates the import of ontological knowledgefrom the OKBC compliant knowledge reposito-ries. Because we are using the task-reductionparadigm, the problem-solving rules of a DISCI-PLE agent are task-reduction rules. DISCIPLE usesan original representation of partially learnedrules called plausible version space (PVS). Muchof the power of the DISCIPLE approach comesfrom this PVS representation of the rules andfrom the multistrategy learning methods usedto learn them. They allow DISCIPLE to incremen-tally learn a rule, starting from only one exam-

ple, and immediately use it in problem solving.The DISCIPLE approach is based on several lev-

els of synergism between the expert who hasthe knowledge to be formalized and the agentthat incorporates knowledge engineeringmethods to formalize it. This multilevelsynergism is achieved through mixed-initiativereasoning that integrates complementaryhuman and automated reasoning to takeadvantage of their respective knowledge, rea-soning styles, and computational strengths.The mixed-initiative reasoning is based on adivision of responsibility between the expertand the agent for those elements of knowledgeengineering for which they have the most apti-tude, such that together they form a completeteam for the development of the agent’s knowl-edge base. From the point of view of the SME,this mixed-initiative approach results in thereplacement of the difficult knowledge engi-neering tasks required to build a knowledgebase, tasks that cannot be performed by anSME, with simpler tasks that can be performedby the expert, as shown in figure 2.

To build a knowledge base, we have first todevelop a model of the application domainthat will make explicit, at a qualitative andinformal level, the way the expert solves prob-lems. In the case of DISCIPLE, we have to modelthe process of solving a specific problem as asequence of qualitative and informal problem-reduction steps, where a complex problem is

Articles

SUMMER 2001 47

Complex knowledge engineering tasks required to build a KB

Simpler tasks that can be performed by a subject matter expert

SME

Cannot

Can

Definedomainmodel

SMESMESMESMESMESME

SMESMESMESMESME

Disciple Approach

Definerules

Verify andupdaterules

Createformal

sentences

Createformal

explanations

Refinedomainmodel

Refineontology

Defineexamples

Critiqueexamples

Understandformal

sentences

Createinformal

hints

Createontology

Figure 2. The General Strategy behind the DISCIPLE Approach.

episode because DISCIPLE will be able to learn arule from this example. Instead of debugging acomplex problem-solving rule, the expertwould only need to critique specific examplesof problem-solving episodes, and DISCIPLE willaccordingly update the corresponding rule.

Most of the time, the expert would not needto create formal sentences or explanations butonly to understand such sentences-explana-tions that were generated by DISCIPLE based oninformal hints provided by the expert.

The Course-of-Action ChallengeProblem

A military COA is a preliminary outline of aplan for how a military unit might attempt toaccomplish a mission. A COA is not a completeplan; it omits many details of the operation,such as exact initial locations of friendly andenemy forces. After receiving orders to plan fora mission, a commander and staff complete adetailed and practiced process of analyzing themission, conceiving and evaluating potentialCOAs, selecting a COA, and preparing detailedplans to accomplish the mission based on theselected COA. The general practice is for thestaff to generate several COAs for a mission andthen make a comparison of these COAs basedon many factors, including the situation, thecommander’s guidance, the principles of war,and the tenets of army operations. The com-mander makes the final decision on whichCOA will be used to generate his/her planbased on the recommendations of the staff andhis/her own experience with the same factorsconsidered by the staff (Jones 1999).

The COA challenge problem consisted ofrapidly developing a knowledge-based cri-tiquing agent that can automatically critiqueCOAs for ground force operations, systemati-cally assess selected aspects of a COA, and sug-gest repairs to it. The role of this agent is to actas an assistant to the military commander,helping the commander to choose betweenseveral COAs under consideration for a certainmission. The agent could also help studentslearn to develop COAs.

The input to the COA critiquing agent con-sists of the description of a COA that includesthe following aspects:

First is the COA sketch, such as the one in thetop part of figure 3. It is a graphic depiction ofthe preliminary plan being considered. Itincludes enough of the high-level structureand maneuver aspects of the plan to show howthe actions of each unit fit together to accom-plish the overall purpose and omits much ofthe execution detail that will be included in

successively reduced to simpler problems.Then we have to build an ontology that willdefine the terms used to express the problemsand their solutions. Finally, we have to defineformal problem-solving rules, verify the rules,and update them. In general, these processesrequire the creation and modification of formalsentences and formal explanations. An SMEcannot perform these tasks. In fact, they arevery hard even for a knowledge engineer, lead-ing to the well-known knowledge-acquisitionbottleneck in the creation of knowledge-basedagents.

In the DISCIPLE approach, the difficult knowl-edge engineering tasks required to develop aknowledge-based agent are replaced with sim-pler tasks that can be performed by an SMEwith limited assistance from a knowledge engi-neer. Instead of developing a complete modelof the application domain, the expert wouldneed to start with an initial model, extendingand refining it with the help of the DISCIPLE

agent. The use of the OKBC knowledge modelfor the DISCIPLE ontology facilitates the importof ontological knowledge from the OKBC com-pliant knowledge repositories. Therefore,instead of creating an ontology from scratch,the expert would only need to update andextend an imported ontology (Boicu et al.1999). Instead of defining a complex problem-solving rule, the expert would only need todefine a specific example of a problem-solving

Articles

48 AI MAGAZINE

Mission:BLUE-BRIGADE2 attacks to penetrate RED-MECH-REGIMENT2 at 130600 Aug in order to enable the completion of seize OBJ-SLAM by BLUE-ARMOR-BRIDADE1.

Close: BLUE-TASK-FORCE1, a balanced task force (MAIN-EFFORT) attacks to penetrate RED-MECH-COMPANY4, then clears RED-TANK-COMPANY2 in order to enable the completion of seize OBJ-SLAM by BLUE-ARMOR-BRIGADE1.

BLUE-TASK-FORCE2, a balanced task force (SUPPORTING-EFFORT1) attacks to fix RED-MECH-COMPANY1 and RED-MECH-COMPANY2 and RED-MECH-COMPANY3 in order to prevent RED-MECH-COMPANY1 and RED-MECH-COMPANY2 and RED-MECH-COMPANY3 from interfering with conducts of the MAIN-EFFORT1,then clears RED-MECH-COMPANY1 and RED-MECH-COMPANY2 and RED-MECH-COMPANY3 and RED-TANK-COMPANY1.

Figure 3. A Sample of a Course-of-Action (COA) Sketch and a Fragment of a COA Statement.

the eventual operational plan. The three pri-mary elements included in a COA sketch are(1) control measures that limit and controlinteractions between units; (2) unit graphicsthat depict known, initial locations and com-position of friendly and enemy units; and (3)mission graphics that depict actions and tasksassigned to friendly units. The COA sketch isdrawn using a palette-based sketching utility.

Second is the COA statement, such as the par-tial one shown in the bottom part of figure 3.It clearly explains what the units in a COA willdo to accomplish the assigned mission. Thistext includes a description of the mission andthe desired end state as well as standard ele-ments that describe purposes, operations,tasks, forms of maneuver, units, and resourcesto be used in the COA. The COA statement isexpressed in a restricted but expressive subsetof English.

Third are selected products of mission analy-sis, such as the areas of operations of the units,avenues of approach, key terrain, unit combatpower, and enemy COAs.

Based on this input, the critiquing agent hasto assess various aspects of the COA, such as itsviability (its suitability, feasibility, acceptabili-ty, and completeness), its correctness (whichconsiders the array of forces, the scheme ofmaneuver, and the command and control),and its strengths and weaknesses with respectto the principles of war and the tenets of armyoperations. The critiquing agent should also beable to clearly justify the assessments made andpropose improvements to the COA.

General Presentation of DISCIPLE-COA

In the HPKB Program, the COA challenge prob-lem was solved by developing an integratedsystem composed of four critiquers, each builtby a different team, to solve a part of the over-all problem: (1) a joint team from Teknowledgeand Cycorp, (2) the EXPECT team from USC/ISI,(3) the LOOM team from USC/ISI, and (4) theDISCIPLE team from George Mason University.All these critiquers shared an input ontologythat contains the terms needed to represent theCOAs.

The COAs to be critiqued were provided byAlphatech. As presented in the previous sec-tion, each such COA is represented by a sketchand a textual description. A statement transla-tor (developed by AIAI of the University ofEdinburgh), a COA sketcher (developed byTeknowledge), and a geographic reasoner(developed by Northwestern University) trans-form and fuse these external representations

into a description in the CYC language, accord-ing to the input ontology. This description isfurther used by all the critiquers.

SAIC completed the integrated system with asolution viewer that provided a uniform pre-sentation of the critiques generated by the fourdeveloped critiquers.

Our critiquer, called DISCIPLE-COA, identifiesthe strengths and the weaknesses of a COAwith respect to the principles of war and thetenets of army operations (U.S. Army 1993).There are nine principles of war: (1) objective,(2) offensive, (3) mass, (4) economy of force, (5)maneuver, (6) unity of command, (7) security,(8) surprise, and (9) simplicity. They providegeneral guidance for the conduct of war at thestrategic, operational, and tactical levels. Thetenets of army operations describe the charac-

Articles

SUMMER 2001 49

Assess COA411 with respect to the Principle of Mass

There is a major strength in COA411 with respect to massbecause BLUE-TASK-FORCE1 is the MAIN-EFFORT1 and itacts on the decisive point of the COA (RED-MECH-COMPANY4) with a force ratio of 10.6, which exceeds a rec-ommended force ratio of 3.0. Additionally, the main effort isassisted by supporting action SUPPRESS-MILITARY-TASK1which also acts on the decisive point. This is good evidenceof the allocation of significantly more than minimum com-bat power required at the decisive point and is indicative ofthe proper application of the principle of mass.

There is a strength in COA411 with respect to mass becauseBLUE-TASK-FORCE1 is the main effort of the COA and it hasbeen allocated 33% of available combat power but this isconsidered just a medium level weighting of the main effort.

There is a strength in COA411 with respect to mass becauseBLUE-MECH-COMPANY8 is a COMPANY-UNIT-DESIGNATION level maneuver unit assigned to be the re-serve. This is considered a strong reserve for a BRIGADE-UNIT-DESIGNATION level COA and would be available tocontinue the operation or exploit success.

Reference: FM 100-5 pg 2-4, KF 113.1, KF 113.2, KF 113.3, KF113.4, KF 113.5 - To mass is to synchronize the effects of allelements of combat power at the proper point and time toachieve decisive results. Observance of the Principle of Massmay be evidenced by allocation to the main effort of signifi-cantly greater combat power than the minimum requiredthroughout its mission, accounting for expected losses. Massis evidenced by the allocation of significantly more thanminimum combat power required at the decisive point.

Figure 4. Solutions Generated by DISCIPLE-COA.

COAs into DISCIPLE’s representation.After the description of a specific COA, such

as the one from figure 3, has been loaded intoDISCIPLE’s knowledge base, the SME can teachthe DISCIPLE agent how to critique it. The expertinvokes the cooperative problem solver, selectsan initial critiquing task (such as “Assess COAwith respect to the Principle of Surprise”), andasks the DISCIPLE agent to solve it. DISCIPLE usesits task-reduction rules to reduce the currenttask to simpler tasks, showing the expert thereductions found. The expert can accept areduction proposed by the agent, can reject it,or can decide to define himself/herself a newreduction, as illustrated in figure 5.

To define a new reduction, the expert usesthe example editor, which, in turn, can invokeseveral knowledge elicitation tools, such as theobject editor, the feature editor, or the task edi-tor, if the specification of the example involvesnew knowledge elements that are not presentin the current knowledge base. Once the reduc-

teristics of successful operations. There are fivesuch tenets: (1) initiative, (2) agility, (3) depth,(4) synchronization, and (5) versatility. Figure4, for example, shows some of the strengths ofthe COA from figure 3 with respect to the prin-ciple of mass, identified by DISCIPLE-COA.

In addition to generating answers in naturallanguage, DISCIPLE also provides the referencematerial based on which answers are generat-ed, as shown in the bottom part of figure 4.Also, the DISCIPLE-COA agent can provide justifi-cations for the generated answers at three lev-els of detail, from a very abstract one (thatshows the general line of reasoning followed)to a very detailed one (that indicates each ofthe knowledge pieces used in generating theanswer).

Figure 5 shows the main modules of DISCIPLE-COA and their interactions. The ontologyimport module was used to integrate the inputCOA ontology into DISCIPLE’s knowledge baseand translate the CYC representation of specific

Articles

50 AI MAGAZINE

ONTOLOGYIMPORT

KNOWLEDGE BASE

RULELEARNER

EXPLANATIONGENERATOR

KNOWLEDGEELICITATION

RULEREFINER

COOPERATIVEPROBLEM SOLVER

EXAMPLEEDITOR

RULEREFINER

prov

ide n

ew

redu

ctio

n

rejectincorrect reduction

accept

correct reduction

TOOLS

task reduction rules

ontology

CYCKNOWLEDGE

BASE

generalizedrule

new objects,features, tasks

newrule

specializedrule

expertreduction

critiquingtask

Figure 5. Expert Agent Interactions.

tion has been defined, the rule learner isinvoked to generalize the example to a task-reduction rule. The rule learner automaticallyinvokes the explanation generator that tries tofind the explanation of why the reductionindicated by the expert is correct. The explana-tion generator proposes several plausible expla-nations from which the expert has to select thecorrect ones. The expert can help the agentfind the correct explanations by providing ahint.

If the expert accepts a reduction proposed bythe agent, then the rule refiner is invoked andcan generalize the rule that has led to thisreduction.

If the expert rejects a reduction proposed bythe agent, then the agent attempts to find anexplanation of why the reduction is not cor-rect, with the explanation generator invoked

again, as described earlier. The explanationfound is used by the rule refiner to specializethe rule.

After a new rule is learned or an existing ruleis refined, the cooperative problem solverresumes the task-reduction process until a solu-tion of the initial problem is found.

In addition to the cooperative problemsolver, DISCIPLE-COA also includes an auton-omous problem solver that is used after DISCI-PLE-COA has been trained.

The next sections describe in more detail thisprocess of developing the disciple-coa agent.

Ontology DevelopmentFor DISCIPLE-COA, an initial ontology wasdefined by importing the input ontology builtby Teknowledge and Cycorp for the COA chal-

Articles

SUMMER 2001 51

<OBJECT>

MILITARY-TASK

GEOGRAPHICAL-REGION

ORGANIZATION

PURPOSE

ACTIONEQUIPMENT

MODERN-MILITARY-ORGANIZATION MILITARY-EQUIPMENT MILITARY-PURPOSE

MILITARY-OPERATION

MILITARY-EVENT

MILITARY-MANEUVER

MECHANIZED-INFANTRY-UNIT-MILITARY-SPECIALTYBLUE-ARMOR-BRIGADE2

AVIATION-UNIT-MILITARY-SPECIALTY

ARMORED-UNIT-MILITARY-SPECIALTYINFANTRY-UNIT-MILITARY-SPECIALTY

ECHELON-OF-UNIT

SOVEREIGN-ALLEGIANCE-OF-ORG

ASSIGNMENT

TASK

TASK

BATALLION-UNIT-DESIGNATION

BLUE-SIDE

MAIN-EFFORT1

CLEAR1

REGULAR-STATUS

BLUE-MECH-COMPANY1

BLUE-MECH-COMPANY2

BLUE-ARMOR-COMPANY1

BLUE-ARMOR-COMPANY2

OPERATIONAL-CONTROL-MILITARY-ORG


TROOP-STRENGTH-OF-UNITOBJECT-ACTED-ON

IS-TASK-OF-OPERATION

PENETRATE1

PENETRATE-MILITARY-TASK

ATTACK2

RED-MECH-COMPANY4

INDICATES-MISSION-TYPE

IS-OFFENSIVE-ACTION-FOR

"military offensiveoperation"

FORCE-RATIO 10.6

RECOMMENDED-FORCE-RATIO3

HAS-SURPRISE-FORCE-RATIO6

COMPLEX-MILITARY-TASK

MILITARY-ATTACK

SUB

CLA

SS-O

F

MODERN-MILITARY-UNIT-DEPLOYABLE

MANEUVER-UNIT-MILITARY-SPECIALTY

INST

ANCE-OF

SUBCLA

SS-O

F

SUBCLASS-OF

BLUE-TASK-FORCE2

PLAN

COA-SPECIFICATION-MICROTHEORY

BLUE-TASK-FORCE1

INST

ANC

E-O

F



…

…

…

……

……

…

…

"military offensiveoperation"

Figure 6. Fragment of the Ontology.

the tasks are themselves represented in the fea-ture hierarchy.

Training the DISCIPLE-COA AgentThe next step in the development of the DISCI-PLE-COA critiquer was to teach DISCIPLE to cri-tique COAs with respect to the principles ofwar and the tenets of army operations. Theexpert loads the description of a specific COA,such as COA411 represented in figure 3, andthen invokes the cooperative problem solverwith a task of critiquing the COA with respectto a certain principle or tenet. DISCIPLE uses itstask-reduction rules to reduce the current taskto simpler tasks, showing the expert the reduc-tions found. The expert can accept a reductionproposed by the agent, reject it, or decide todefine a new reduction. From each such inter-action, DISCIPLE will either refine a previouslylearned rule or will learn a new task-reductionrule. After a new rule is learned or an existingrule is refined, the cooperative problem solverresumes the task-reduction process until a solu-tion of the initial problem is found.

Initially, DISCIPLE does not contain any rules.Therefore, all the problem-solving steps (thatis, task reductions) must be provided by theexpert, as illustrated in figure 7 and explainedin the following.

To assess COA411 with respect to the princi-ple of security, the expert and DISCIPLE need acertain amount of information, which isobtained by asking a series of questions. Theanswer to each question allows one to reducethe current assessment task to a more detailedone. This process continues until the expertand DISCIPLE have enough information aboutCOA411 to make the assessment. As shown infigure 7, the initial task is reduced to that ofassessing the security of COA411 with respectto the countering of enemy reconnaissance.Then one asks whether there is any enemyreconnaissance unit present in COA411. Theanswer identifies RED-CSOP1 as such a unitbecause it is performing the task SCREEN1.Therefore, the task of assessing security forCOA411 with respect to countering enemyreconnaissance is now reduced to the better-defined task of assessing security when enemyreconnaissance is present. The next question toask is whether the enemy reconnaissance unitis destroyed. In the case of COA411, RED-CSOP1is destroyed by the task DESTROY1. Therefore,one can conclude that there is a strength inCOA411 with respect to the principle of secu-rity because the enemy reconnaissance unit iscountered.

Figure 8 illustrates the process of teaching

lenge problem. The imported ontology was fur-ther developed by using the ontology-buildingtools of DISCIPLE. These tools include specializedbrowsers and editors for the various knowledgepieces of DISCIPLE (for example, the object edi-tor, the object feature editor, the task featureeditor, the hierarchy browser, and the associa-tion browser).

DISCIPLE’s ontology includes objects, features,and tasks, all represented as frames, accordingto the knowledge model of the OKBC protocol(Chaudhri et al. 1998).

The objects represent either specific individ-uals or sets of individuals. The objects are hier-archically organized according to the general-ization relation (subclass-of/superclass-of andinstance-of/type-of). Figure 6, for example, pre-sents a fragment of the object ontology. Thetop part represents the upper level of the objectontology that identifies the types of conceptrepresented in the ontology. They include GEO-GRAPHICAL-REGION, ORGANIZATION, EQUIPMENT, andACTION. Each of these concepts is the top of aspecialized hierarchy, such as the hierarchy oforganizations showed in the left part of figure6. The leaves of this hierarchy are specific mil-itary units, corresponding to a specific COA tobe critiqued by DISCIPLE. Each concept andinstance of the object hierarchy is described byspecific features and values. For example, thebottom part of figure 6 shows the descriptionof the specific military unit called BLUE-TASK-FORCE1. BLUE-TASK-FORCE1 is described as both anARMORED-UNIT-MILITARY-SPECIALTY and a MECHA-NIZED-INFANTRY-UNIT-MILITARY-SPECIALTY. The otherfeatures describe BLUE-TASK-FORCE1 as being atthe battalion level; belonging to the blue side,designated as the main effort of the blue side;performing two tasks, PENETRATE1 and CLEAR1;having a regular strength; and having under itsoperational control four other units. The val-ues of the features of BLUE-TASK-FORCE1 are them-selves described in the same way. For example,one of the tasks performed by BLUE-TASK-FORCE1is PENETRATE1. PENETRATE1 is defined as a penetra-tion task and, therefore, inherits all the featuresof the penetration tasks in addition to the fea-tures that are directly associated with it.

The hierarchy of objects is used as a general-ization hierarchy for learning by the DISCIPLE

agent. One way to generalize an expression isto replace an object with a more general onefrom such a hierarchy. For example, PENETRATE-MILITARY-TASK from the bottom right side of fig-ure 6 can be generalized to COMPLEX-MILITARY-TASK, MILITARY-MANEUVER, or MILITARY-ATTACK. Thegoal of the learning process is to select the cor-rect generalization.

The features used to describe the objects and

Articles

52 AI MAGAZINE

DISCIPLE. The left side represents the reasoningprocess of the expert, where the question andthe answer are in free natural language format.Although this line of reasoning is natural to ahuman expert, a learning agent cannot under-stand it. The explanation that would be under-stood by the agent is represented in the upperright part of figure 8 and consists of variousrelations between certain elements from itsontology. The first explanation piece states, inDISCIPLE’s formal language, that RED-CSOP1 is anenemy unit. The second explanation pieceexpresses the fact that RED-CSOP1 is performingthe action SCREEN1. Finally, the last explanationpiece expresses the fact that SCREEN1 is a recon-naissance action. Although an expert canunderstand the meaning of these formalexpressions, he/she cannot define thembecause he/she is not a knowledge engineer.

For one thing, he/she would need to use theformal language of the agent. Moreover, he/shewould also need to know the names of thepotentially many thousands of concepts andfeatures from the agent’s ontology.

Although defining the formal explanationsof this task-reduction step is beyond the indi-vidual capabilities of the expert and the agent,it is not beyond their joint capabilities. Findingthese explanation pieces is a mixed-initiativeprocess of searching the agent’s ontologybecause an explanation piece is a path ofobjects and relations in this ontology. Inessence, the agent will use analogical reasoningand help from the expert to identify and pro-pose a set of plausible explanation pieces fromwhich the expert will have to select the correctones. One explanation-generation strategy isbased on an ordered set of heuristics for analog-

Articles

SUMMER 2001 53

ASSESS-SECURITY-WRT-COUNTERING-ENEMY-RECONAISSANCEFOR-COA COA411

ASSESS-SECURITY-WHEN-ENEMY-RECON-IS-PRESENTFOR-COA COA411FOR-UNIT RED-CSOP1FOR-RECON-ACTION SCREEN1

ASSESS-COA-WRT-PRINCIPLE-OF-SECURITYFOR-COA COA411 R

$AC

WPO

S-001R

$ASW

CER

-001

R$A

SWER

IP-001

Does the COA include security and counter-recon actions, a security element, a rear element, and identify risks?

I consider enemy reconnaissance

Is an enemy reconnaissance unit present?

Yes, RED-CSOP1 which is performingthe reconnaissance action SCREEN1

Is the enemy reconnaissance unit destroyed?

RuleLearning

RuleLearning

RuleLearning

Yes, RED-CSOP1 is destroyed by DESTROY1

REPORT-STRENGTH-IN-SECURITY-BECAUSE-OF-COUNTERING-ENEMY-RECON

FOR-COA COA411FOR-UNIT RED-CSOP1FOR-RECON-ACTION SCREEN1

DESTROY1"high"

FOR ACTIONWITH-IMPORTANCE

”

Figure 7. Task Reductions Indicated by the Expert.

exact applicability condition that DISCIPLE

attempts to learn. Initially, the plausible lower-bound condition covers only the example infigure 8, restricting the variables from the ruleto take only the values from this example. Italso includes the relations between these vari-ables that have been identified as relevant inthe explanation of the example. The plausibleupper-bound condition is the most generalgeneralization of the plausible lower-boundcondition. It is obtained by taking into accountthe domains and the ranges of the featuresfrom the plausible lower-bound conditions andthe tasks to determine the possible values ofthe variables. The domain of a feature is the setof objects that might have this feature. Therange is the set of possible values of this fea-ture. For example, ?O2 is the value of the taskfeature FOR-UNIT and has as features SOVEREIGN-

ical reasoning. These heuristics exploit thehierarchies of objects, features, and tasks toidentify the rules that are similar to the currentreduction and use their explanations as a guideto search for similar explanations of the cur-rent example. This cooperative explanation-generation process proved to be effective, asdemonstrated by the successful knowledge-acquisition experiment described in this arti-cle.

From the example reduction and its explana-tion in figure 8, DISCIPLE automatically generat-ed the PVS rule in figure 9, which is an if-thenrule, the components of which are generaliza-tions of the elements of the example in figure8. In addition, the rule contains two conditionsfor its applicability: (1) a plausible lower-boundcondition and (2) a plausible upper-boundcondition. These conditions approximate an

Articles

54 AI MAGAZINE

Figure 8. Teaching DISCIPLE to Reduce a Task.

Answer:

Yes, RED-CSOP1 which isperforming the reconnaissance

action SCREEN1.

Question:Is an enemy reconnaissance

unit present?

Learn the rule



Logic

Explanation:RED-CSOP1 SOVEREIGN-ALLEGIANCE-OF-ORG RED-SIDERED-CSOP1 TASK SCREEN1SCREEN1 IS INTELLIGENCE-COLLECTION-MILITARY-TASK

Natural Language

Rule: R$ASWCER-001

Plausible Upper Bound Condition?O1 IS COA-SPECIFICATION-MICROTHEORY?O2 IS MODERN-MILITARY-UNIT-DEPLOYABLE SOVEREIGN-ALLEGIANCE-OF-ORG ?O4 TASK ?O3?O3 IS INTELLIGENCE-COLLECTION-MILITARY-TASK?O4 IS RED-SIDE

IF the task to accomplish isASSES-SECURITY-WRT-COUNTERING-ENEMY-RECONAISSANCE FOR-COA O1?

Explanation ?O2 SOVEREIGN-ALLEGIANCE-OF-ORG ?O4 IS RED-SIDE ?O2 TASK ?O3 IS INTELLIGENCE-COLLECTION-MILITARY-TASK

THEN accomplish the taskASSES-SECURITY-WHEN-ENEMY-RECON-IS-PRESENT FOR-COA ?O1 FOR-UNIT ?O2 FOR-RECON-ACTION ?O3

Plausible Lower Bound Condition ?O1 IS COA411 ?O2 IS RED-CSOP1 SOVEREIGN-ALLEGIANCE-OF-ORG ?O4 TASK ?O3 ?O3 IS SCREEN1 ?O4 IS RED-SIDE

Mai

n C

ondi

tion

Question: Is an enemy reconnaissanceunit present?

Answer: Yes, ?O2 which is performing the reconnaissance action?O3.

{}

}

ALLEGIANCE-OF-ORG and TASK. Therefore, any val-ue of ?O2 has to be in the intersection of therange of FOR-UNIT, the domain of SOVEREIGN-ALLE-GIANCE-OF-ORG, and the domain of TASK. Thisintersection is MODERN-MILITARY-UNIT-DEPLOYABLE.

The learned PVS rules, such as the one in fig-ure 9, are used in problem solving to generatetask reductions with different degrees of plau-sibility, depending on which of their condi-tions are satisfied. If the plausible lower-boundcondition is satisfied, then the reduction islikely to be correct. If the plausible lower-bound condition is not satisfied, but the plau-sible upper-bound condition is satisfied, thenthe solution is considered only plausible. Anyapplication of a PVS rule however, successful ornot, provides an additional (positive or nega-tive) example, and possibly an additionalexplanation, that is used by the agent to fur-ther improve the rule through the generaliza-tion or specialization of its conditions.

Let us consider again the specific task reduc-tions from figure 7. At least for the elementarytasks, such as the one from the bottom of thefigure, the expert also needs to express them innatural language:

There is a strength with respect to surprisein COA411 because it contains aggressivesecurity-counterreconnaissance plans,destroying enemy intelligence collectionunits and activities. Intelligence collectionby RED-CSOP1 will be disrupted by itsdestruction by DESTROY1.

Similarly, the expert would need to indicate thesource material for the concluded assessment.The learned rules contain generalizations ofthese phrases that are used to generate answersin natural language, as illustrated in figure 4.Similarly, the generalizations of the questionsand the answers from the rules applied to gen-erate a solution are used to produce an abstractjustification of the reasoning process.

As DISCIPLE-COA learns PVS rules, it can usethem to propose routine or innovative solu-tions to the current problems. The routine solu-tions are those that satisfy the plausible lower-bound conditions of the rules and are likely tobe correct. Those that are not correct are keptas exceptions to the rule. The innovative solu-tions are those that do not satisfy the plausiblelower-bound conditions but satisfy the plausi-ble upper-bound conditions. These solutionsmight or might not be correct, but in each case,they will lead to a refinement of the rules thatgenerated them. Let us consider the situationillustrated in figure 10. After it has been shownhow to critique COA411 with respect to theprinciple of security, DISCIPLE is asked to critiqueCOA421. COA421 is similar to COA411 except

that in this case, the enemy recon unit is notdestroyed. Because of this similarity, DISCIPLE-COA is able to propose the two top reductions infigure 10. Both of them are innovative reduc-tions that are accepted by the expert. There-fore, DISCIPLE-COA generalizes the plausible low-er-bound conditions of the correspondingrules, as little as possible, to cover these reduc-tions and remain less general or, at most, asgeneral as the corresponding plausible upper-bound conditions.

The last reduction step in figure 10 has to beindicated by the expert because no rule of DIS-CIPLE-COA is applicable. We call the expert-pro-vided reduction a creative problem-solvingstep. From each such reduction, DISCIPLE-COA

learns a new task-reduction rule, as was illus-trated earlier.

Articles

SUMMER 2001 55

Rule: R$ASWCER-001

Plausible Upper Bound Condition?O1 IS COA-SPECIFICATION-MICROTHEORY?O2 IS MODERN-MILITARY-UNIT-DEPLOYABLE SOVEREIGN-ALLEGIANCE-OF-ORG ?O4 TASK ?O3?O3 IS INTELLIGENCE-COLLECTION-MILITARY-TASK?O4 IS RED-SIDE

IF the task to accomplish isASSES-SECURITY-WRT-COUNTERING-ENEMY-RECONAISSANCE FOR-COA O1?

Explanation ?O2 SOVEREIGN-ALLEGIANCE-OF-ORG ?O4 IS RED-SIDE ?O2 TASK ?O3 IS INTELLIGENCE-COLLECTION-MILITARY-TASK

THEN accomplish the taskASSES-SECURITY-WHEN-ENEMY-RECON-IS-PRESENT FOR-COA ?O1 FOR-UNIT ?O2 FOR-RECON-ACTION ?O3

Plausible Lower Bound Condition ?O1 IS COA411 ?O2 IS RED-CSOP1 SOVEREIGN-ALLEGIANCE-OF-ORG ?O4 TASK ?O3 ?O3 IS SCREEN1 ?O4 IS RED-SIDE

Mai

n C

ondi

tion

Question: Is an enemy reconnaissanceunit present?

Answer: Yes, ?O2 which is performing the reconnaissance action?O3.

Figure 9. Plausible Version-Space Rule Learned from the Example and the Explanation in Figure 8.

ure 9. With disciple, the domain expert needsonly to define an example reduction becausedisciple learns and refines the correspondingrule. This approach works well, as is demon-strated by the intense experimental studiesconducted with disciple that are reported inthe next section.

Evaluation of the Course-of-Action Critiquers

In addition to George Mason University, otherthree research groups have developed COA cri-tiquers as part of the HPKB Program. Teknowl-edge and Cycorp have developed a critiquerbased on the CYC system (Lenat 1995). The oth-er two critiquers have been developed atUSC/ISI, one based on the EXPECT system (Kim

Through refinement, the task-reduction rulescan become significantly more complex thanthe rule in figure 9 (Boicu et al. 2000). For exam-ple, when a reduction proposed by DISCIPLE isrejected by the expert, DISCIPLE attempts to findan explanation for why the reduction is wrong.Then the rule can be refined with an except-when plausible version-space condition. Thebounds of this version space are generalizationsof the explanations that should not hold for thereduction rule to be applicable.

In any case, comparing the left-hand side offigure 8 (which is defined by the domainexpert) with the rule from figure 9 (which islearned by disciple) suggests the usefulness ofdisciple for knowledge acquisition. In the tradi-tional knowledge engineering approach, aknowledge engineer would need to manuallydefine and debug a rule such as the one in fig-

Articles

56 AI MAGAZINE



REPORT-WEAKNESS-IN-SECURITY-BECAUSE-ENEMYRECON-IS-NOT-COUNTERED

FOR-COA COA421FOR-UNIT RED-CSOP2FOR-RECON-ACTION SCREEN2WITH-IMPORTANCE “high ”

ASSESS-COA-WRT-PRINCIPLE-OF-SECURITYFOR-COA COA421 R

$AC

WPO

S-001R

$ASW

CER

-001

R$A

SWER

IP-002Does the COA include security and counter-recon actions,

a security element, a rear element, and identify risks?

I consider enemy reconnaissance

Is an enemy reconnaissance unit present?

Yes, RED-CSOP2 which is performingthe reconnaissance action SCREEN2

No

Is the enemy reconnaissance unit destroyed?

RuleLearning

RuleRefinement

RuleRefinement

Figure 10. Cooperative Problem Solving and Learning.

and Gil 1999) and the other based on the LOOM

system (MacGregor 1999).1 All the critiquerswere evaluated as part of the HPKB’s annualevaluation that took place from 6 to 16 July1999 and included five evaluation items ofincreasing difficulty. Each item consisted ofdescriptions of various COAs and a set of ques-tions to be answered about each of them. Item1 consisted of COAs and questions that werepreviously provided by DARPA to guide thedevelopment of the COA critiquing agents.Item 2 included new test questions about thesame COAs. Items 3, 4, and 5 consisted of newCOAs that were increasingly more complexand required further development of the COAagents to properly answer the asked questions.Each of items 3, 4, and 5 consisted of two phas-es. In the first phase, each team had to provideinitial system responses. Then the evaluatorissued the model answers, and each team hada limited amount of time to repair its system,perform further knowledge acquisition, andgenerate revised system responses.

The responses of each system were scoredby a team of domain experts along the follow-ing dimensions and associated weights: cor-rectness—50 percent (matches model answer

or is otherwise judged to be correct), justifica-tion—30 percent (scored on presence, sound-ness, and level of detail), lay intelligibility—10percent (degree to which a lay observer canunderstand the answer and the justification),sources—10 percent (degree to which appro-priate sources are noted), and proactivity—10-percent extra credit (appropriate correctiveactions or other information suggested toaddress the critique). Based on these scores,several classes of metrics have been computed,including recall and precision. Recall isobtained by dividing the score for all answersprovided by a critiquer to the total number ofmodel answers for the asked questions. Thisnumber was over 100 percent with our criti-quer, primarily because of the extra creditreceived for generating additional critiquesthat were not among the model answers pro-vided by the evaluator. Precision is obtained bydividing the same score by the total numberof answers provided by the system (both themodel answers provided by the evaluator andthe new answers provided by the critiquer).The results obtained by the four evaluated cri-tiquers are presented in figure 11. Thesegraphs also show the averages of these results,

Articles

SUMMER 2001 57

Metric: Recall (Total Score)

56.81%63.71%

114.69%

70.20%

0.00%

20.00%

40.00%

60.00%

80.00%

100.00%

120.00%Recall

Tek/Cyc ISI-Expect GMU ISI-Loom

84.20%

ALL

Tek/Cyc ISI-Expect GMU ISI-Loom ALLMetric: Precision (Total Score)

62.61%

76.01%81.99%

57.48%

0.00%

20.00%

40.00%

60.00%

80.00%

100.00%

Recall

Tek/Cyc ISI-Expect GMU ISI-Loom

75.01%

ALL

0.00%

20.00%

40.00%

60.00%

80.00%

100.00%

120.00%

Correctness ProactivityJustification Intelligibility Source

RecallRecall Breakdown by Criteria

Total

Precision

IntelligibilityCorrectness Justification Source Proactivity0.00%

10.00%

20.00%

30.00%40.00%

50.00%

60.00%70.00%

80.00%

90.00%100.00%

Precision Breakdown by Criteria

Total

Figure 11. The Performance of the Developed Course-of-Action Critiquers and the Integrated System.

and had three phases: (1) joint training (days 1to 3), (2) an individual teaching experiment(day 4), and (3) a joint discussion of the exper-iment (day 5). The entire experiment wasvideotaped. The training for the experimentincluded a detailed presentation of DISCIPLE’sknowledge representation, problem-solvingand learning methods, and tools. For the teach-ing experiment, each expert received a copy ofDISCIPLE-COA with a partial knowledge base. Thisknowledge base was obtained by removing thetasks and the rules from the complete knowl-edge base of DISCIPLE-COA. That is, the knowl-edge base contained the complete ontology ofobjects, object features, and task features. Wealso provided the experts with the descriptionsof three COAs—(1) COA411, (2) COA421, and(3) COA51—to be used for training DISCIPLE.These were the COAs used in the final phasesof DARPA’s evaluation of all the critiquers.Finally, we provided and discussed with theexperts the modeling of critiquing of theseCOAs with respect to the principles of offen-sive and security. That is, we provided theexperts with specific task reductions such asthe one from figure 7 to guide them in teach-ing DISCIPLE-COA. Each expert then taught DISCI-PLE-COA independently while he/she was super-vised by a knowledge engineer, whose role wasto help the expert if he/she reached an impassewhile he/she was using DISCIPLE.

Figure 13 shows the evolution of the knowl-edge base during the teaching process for oneof the experts, results that are representativefor all four experts. In the morning the experttaught DISCIPLE to critique COAs with respect tothe principle of offensive and in the afternoonhe taught it to critique COAs with respect tothe principle of security. In both cases, theexpert first used COA411, then COA422 andthen COA51. As one can see from figure 13, DIS-CIPLE initially learned more rules, and then theemphasis shifted to rule refinement. Therefore,the increase in the size of the knowledge baseis greater toward the beginning of the trainingprocess for each principle. The teaching for theprinciple of offensive took 101 minutes. Dur-ing this time, DISCIPLE learned 14 tasks and 14rules (147 simple axioms equivalent). Theteaching for security took place in the after-noon and consisted of 72 minutes of expert-DISCIPLE interactions. During this time, DISCIPLE

learned 14 tasks and 12 rules (136 simpleaxioms equivalent). There was no or very lim-ited assistance from the knowledge engineerwith respect to teaching. The knowledge-acqui-sition rate obtained during the experiment wasvery high (9 tasks and 8 rules an hour, or 98simple axioms equivalent an hour). At the end

which represent the recall and the precision ofthe integrated system.

Figure 12 compares the recall and the cover-age of the developed critiquers for the last threemost complex items of the evaluation. For eachitem, the beginning of each arrow shows thecoverage and recall for the initial testing phase,and the end of the arrow shows the same datafor the modification phase. In this graph, theresults that are above and to the right are supe-rior to the other results. This graph also showsthat all the systems increased their coverageduring the evaluation. In particular, the knowl-edge base of DISCIPLE increased by 46 percent(from the equivalent of 6229 simple axioms to9092 simple axioms), which represents a veryhigh rate of knowledge acquisition of 286 sim-ple axioms a day.

Direct Knowledge Acquisitionfrom Subject-Matter Experts

During August 1999, we conducted a one-weekknowledge-acquisition experiment with DISCI-PLE-COA at the U.S. Army Battle Command Bat-tle Lab in Fort Leavenworth, Kansas, to test theclaim that domain experts that do not haveprior knowledge engineering experience canteach DISCIPLE-COA (Tecuci et al. 2000). Theexperiment involved four such military experts

Articles

58 AI MAGAZINE

Rec

all

0

20

40

60

80

100

120

140

160

50% 100%

25% 75%

3

5

4

4

5

3

3

4

5

Coverage

(Evaluation Items 3, 4, and 5)

TEK/CYC

GMU (Disciple)

ISI(Expect)

ISI(Expect)

ISI(Expect)

GMU(Disciple)

GMU (Disciple)TEK/CYC

TEK/CYC

Figure 12. Coverage versus Recall, Prerepair, and Postrepair.

of this training process, DISCIPLE-COA was able tocorrectly identify 17 strengths and weaknessesof the 3 COAs with respect to the principles ofoffensive and security.

After the experiment, each expert was askedto fill out a detailed questionnaire designed tocollect subjective data for usability evaluation.All the answers took into account that DISCIPLE-COA was a research prototype and not a com-mercial product and were rated based on a scaleof agreement with the question from 1 to 5,with 1 denoting not at all and 5 denoting very.For illustration, table 1 shows three questionsand the answers provided by the four experts.

ConclusionsWe introduced the concept of a learning agentshell and a methodology for rapid develop-ment of knowledge bases and agents based onthe DISCIPLE learning agent shell. The DISCIPLE

shell and methodology have been applied tothe development of a critiquing agent that actsas an assistant to a military commander. Thisapproach and the developed agent have beenevaluated in two intensive studies. The firststudy concentrated on the quality of the devel-oped critiquer and the ability to rapidly extendit by its developers and SMEs. The second studyconcentrated on the ability of domain expertsto extend the knowledge base of the critiquer

with limited assistance from knowledge engi-neers. Both studies have shown that DISCIPLE

has reached a significant level of maturity,being usable to rapidly develop complexknowledge-based agents.

The two main factors that contributed to thesuccess of DISCIPLE-COA are (1) the synergisticcollaboration between the SME and DISCIPLE indeveloping the knowledge base and (2) themultistrategy learning method of DISCIPLE thatis based on the plausible version-space repre-sentation. Our research on PVSs has its originsin Mitchell’s (1997) influential work on versionspaces and his candidate elimination algo-rithm. We extended them along several dimen-sions, which led to a powerful and practicalmixed-initiative multistrategy learning ap-proach that synergistically integrates a widerange of knowledge-acquisition and machinelearning strategies, including apprenticeshiplearning, empirical inductive learning fromexamples and explanations, and analogicallearning. This method is based on a powerfulknowledge representation language that in-cludes the frame-based OKBC knowledge mod-el for the representation of the ontologicalknowledge and complex task-reduction ruleswith multiple conditions. Moreover, we do notmake the assumption that the representationspace for learning needs to be completelydefined before learning can take place. On the

Articles

SUMMER 2001 59

0

50

100

150

200

250

300

350

400

Initial 411

Training for Offensive

411 51 411

Task + Rule axioms

Rule axioms

Task axioms

Nu

mbe

r of

Axi

oms

421 51 411 421421 51

Training for Security

421 51

Figure 13. The Evolution of the Knowledge Base during the Knowledge-Acquisition Experiment.

oratory, Air Force Material Command, USAF,under agreement F30602-00-2-0546, grantF49620-97-1-0188 and grant F49620-00-1-0072. The U.S. government is authorized toreproduce and distribute reprints for govern-mental purposes notwithstanding any copy-right annotation thereon. The views and con-clusions contained herein are those of theauthors and should not be interpreted as nec-essarily representing the official policies orendorsements, either expressed or implied, ofDARPA, AFOSR, the Air Force Research Labora-tory, or the U.S. government. The evaluation ofthe COA critiquers in the HPKB program wasconducted by Alphatech. The experts that par-ticipated in the BCBL knowledge-acquisitionexperiment were Lt. Col. John N. Duquette, Lt.Col. Jay E. Farwell, Maj. Michael P. Bowman,and Maj. Dwayne E. Ptaschek. Florin Ciucu,Cristian Levcovici, Cristina Cascaval, PingShyr, and Kathryn Wright contributed to DISCI-PLE-COA.

Notes1. R. MacGregor, 1999. Retrospective on LOOM.Available online at www.isi.edu/isd/ LOOM/papers/macgregor/Loom_Retrospective.html.

ReferencesBoicu, M.; Wright, K.; Marcu, D.; Lee, S. W.; Bowman,M.; and Tecuci, G. 1999. The DISCIPLE Integrated Shelland Methodology for Rapid Development of Knowl-

contrary, the representation language is as-sumed to be incomplete and partially incorrectand is itself evolving during rule learningthrough the improvement of the ontology.Because the learning process takes place in anevolving representation language, the variousplausible bounds of a rule are subject to heuris-tic transformations that involve both general-ization and specialization operations for eachbound. These transformations are guided byhints, explanations, and analogies. Therefore,the learning process is efficient and does notsuffer from any combinatorial explosion. Also,learning can take place even in the presence ofexceptions when there is no rule that discrim-inates between the positive examples and thenegative examples.

To conclude, our long-term vision for DISCIPLE,which guides our future work, is to evolve it toa point where it will allow normal computerusers to build and maintain knowledge basesand knowledge-based agents as easily as theynow use personal computers for text processing.

AcknowledgmentsThis research was done in the George MasonUniversity Learning Agents Laboratory (LAL-AB). Research of the LALAB is sponsored by theDefense Advanced Research Projects Agency(DARPA), the Air Force Office of ScientificResearch (AFOSR), and Air Force Research Lab-

Articles

60 AI MAGAZINE

Table 1. Sample Questions Answered by the Experts.

Do you think that Discipleis a useful tool forknowledge acquisition?

Do you think that Discipleis a useful tool for ProblemSolving?

Were the procedures /processes used in Disciplecompatible with Armydoctrine and/or decisionmaking processes?

Questions Answers

• Rating 5. Absolutely! The potential use of this tool by domain experts is only limited by their imagination—not their AI programming skills.

• 5 • 4

• Yes, it allowed me to be consistent with logical thought.

• Rating 5. Yes.• 5 (absolutely)• 4• Yes. As it develops and becomes tailored to the user, it will simplify the tedious tasks.

• Rating 5. As a minimum yes, as a maximum—better!• This again was done very well.• 4 • 4

edge-Based Agents. In Proceedings of theSixteenth National Conference on ArtificialIntelligence, 900–901. Menlo Park, Calif:American Association for Artificial Intelli-gence.

Boicu, M.; Tecuci, G.; Marcu, D.; Bowman,M.; Shyr, P.; Ciucu, F.; and Levcovici, C.2000. DISCIPLE-COA: From Agent Program-ming to Agent Teaching. In Proceedings ofthe Seventeenth International Conference onMachine Learning, 73–80. San Francisco,Calif.: Morgan Kaufmann.

Chaudhri, V. K.; Farquhar, A.; Fikes, R.;Karp, P. D.; and Rice, J. P. 1998. OKBC: AProgrammatic Foundation for KnowledgeBase Interoperability. In Proceedings of theFifteenth National Conference on ArtificialIntelligence, 600–607. Menlo Park, Calif.:American Association for Artificial Intelli-gence.

Clancey, W. J. 1984. NEOMYCIN: Reconfigur-ing a Rule-Based System with Applicationto Teaching. In Readings in Medical ArtificialIntelligence, eds. W. J. Clancey and E. H.Shortliffe, 361–381. Reading, Mass.: Addi-son-Wesley.

Cohen, P.; Schrag, R.; Jones, E.; Pease, A.;Lin, A.; Starr, B.; Gunning, D.; and Burke,M. 1998. The DARPA High-PerformanceKnowledge Bases Project. AI Magazine19(4): 25–49.

Fikes, R.; Farquhar, A.; and Rice, J. 1997.Tools for Assembling Modular Ontologiesin ONTOLINGUA. In Proceedings of the Four-teenth National Conference on ArtificialIntelligence, 436–441. Menlo Park, Calif.:American Association for Artificial Intelli-gence.

Jones, E. 1999. HPKB Course of ActionChallenge Problem Specification, Alphate-ch, Inc., Burlington, Massachusetts.

Kim, J., and Gil, Y. 1999. Deriving Expecta-tions to Guide Knowledge Base Creation. InProceedings of the Sixteenth National Con-ference on Artificial Intelligence, 235–241,Menlo Park, Calif.: American Associationfor Artificial Intelligence.

Lenat, D. B. 1995. CYC: A Large-Scale Invest-ment in Knowledge Infrastructure. Commu-nications of the ACM 38(11): 33–38.

MacGregor, R. 1991. The Evolving Technol-ogy of Classification-Based Knowledge Rep-resentation Systems. In Principles of SemanticNetworks: Explorations in the Representationsof Knowledge, ed. J. Sowa, 385–400. San Fran-cisco, Calif.: Morgan Kaufmann.

Mitchell, T. M. 1997. Machine Learning. NewYork: McGraw-Hill.

Tecuci, G. 1998. Building Intelligent Agents:An Apprenticeship Multistrategy Learning The-ory, Methodology, Tool, and Case Studies. SanDiego, Calif.: Academic.

Tecuci, G.; Boicu, M.; Bowman, M.; Marcu,

D.; Shyr, P.; and Cascaval, C. 2000. AnExperiment in Agent Teaching by SubjectMatter Experts. International Journal ofHuman-Computer Studies 53(10): 583–610.

Tecuci, G.; Boicu, M.; Wright, K.; Lee, S. W.;Marcu, D.; and Bowman, M. 1999. An Inte-grated Shell and Methodology for RapidDevelopment of Knowledge-Based Agents.In Proceedings of the Sixteenth NationalConference on Artificial Intelligence,250–257. Menlo Park, Calif.: AmericanAssociation for Artificial Intelligence.

U.S. Army. 1993. U.S. Army Field Manual,100-5, Operations, Headquarters, Depart-ment of the Army, Washington, D.C.

Gheorghe Tecuci is pro-fessor of computer sci-ence and director of theLearning Agents Labora-tory at George MasonUniversity and a mem-ber of the RomanianAcademy. He receivedtwo Ph.D.s in computer

science, one from the University of Paris-South, Orsay, France, and the other fromthe Polytechnic University of Bucharest,Romania. He has published over 100 scien-tific papers and 5 books, including BuildingIntelligent Agents: An Apprenticeship Multi-strategy Learning Theory, Methodology, Tool,and Case Studies, Machine Learning: A Multi-strategy Approach, and Machine Learning andKnowledge Acquisition: Integrated Approaches.His e-mail address is [email protected].

Mihai Boicu is a Ph.D.student in the School ofInformation Technolo-gy and Engineering atGeorge Mason Universi-ty and a research assis-tant in the LearningAgents Laboratory. Pre-viously he was a tenured

teacher in the Computer Science Depart-ment of the National School of ComputerScience, Bucharest, Romania. His researchinterests include AI, machine learning,intelligent agents, knowledge acquisition,knowledge representation, problem solv-ing, theory of algorithms, and education.He is a member of the American Associa-tion for Artificial Intelligence. His e-mailaddress is [email protected].

Lt.Col. Michael Bowman is an active dutyArmy officer and a Ph.D. candidate atGeorge Mason University. He received aB.S. from Ouachita Baptist University and

an M.S. from the NavalPostgraduate School. Hewas the U.S. Army prod-uct manager for Com-munications and Intelli-gence Support Systemsand has had a variety ofacquisition, automation,and tactical assignments

at the Defense Intelligence Agency and theU.S. Military Academy and in several ArmyField Artillery battalions. His e-mail addressis [email protected].

Dorin Marcu is a Ph.D.candidate in computerscience and a researchassistant in the LearningAgents Laboratory atGeorge Mason Universi-ty. His current researchinterests are intelligentuser interfaces and

machine learning. He is a member of theAmerican Association for Artificial Intelli-gence. His e-mail address is [email protected].

Murray Burke hasserved as the programmanager of the High-Performance KnowledgeBase Project and is cur-rently the program man-ager of the Rapid Knowl-edge Formation Projectat the Defense Advanced

Research Projects Agency. He has over 30years in command and control informationsystems research and development. He wascofounder and executive vice president ofKnowledge Systems Concepts, Inc., a soft-ware company specializing in defenseapplications of advanced information tech-nologies. Burke is a member of the Ameri-can Association for Artificial Intelligence.

Articles

SUMMER 2001 61

Articles

62 AI MAGAZINE

P a u l J . F e l t o v i c h

K e n n e t h M . F o r d

R o b e r t R . H o f f m a n

Expertise in Context:Human and Machine

COMPUTERIZED “expertsystems” are among

the best knownapplications of artificial intel-ligence. But what is expertise?The nature of knowledge and

expertise, and their relation tocontext, is the focus of active

discussion — even controversy— among psychologists,

philosophers, computer scien-tists, and other cognitive sci-

entists. The questions reach tothe very foundations of cogni-

tive theory — with new per-spectives contributed by the

social sciences. These debatesabout the status and nature ofexpert knowledge are of inter-

est to and informed by theartificial intelligence commu-nity — with new perspectives

contributed by “construc-tivists” and “situationalists.”

The twenty-three essays inthis volume discuss the essen-

tial nature of expert knowl-edge, as well as such questions

such as how “expertise” dif-fers from mere “knowledge,”the relation between the indi-

vidual and group processesinvolved in knowledge in gen-eral and expertise in particu-lar, the social and other con-

texts of expertise, howexpertise can be assessed, and

the relation between humanand computer expertise.

The cover is a reproduction of “C

lown T

hree” by Max Papart. It is reproduced w

ith per-m

ission, courtesy of Nahan G

alleries, New

York, New

York.

690 pp., ISBN 0-262-56110-7

Published by the AAAI Press / The MIT Press

To order call 800-356-0343 (US and Canada)or (617) 625-8569. Distributed by The MIT Press, 5 Cambridge Center, Cambridge, MA 02142

AI Magazine Volume 22 Number 2 (2001) (© AAAI) An ...

Documents