From: AAAI Technical Report WS-97-05. Compilation copyright © 1997, AAAI (www.aaai.org). All rights reserved. Constraints and Agents in MADEsmart Leo Obrst Boeing Defense & Space Group Helicopters Division Advanced Computing Technologies P.O. Box 16858, MSP29-99 Philadelphia, PA19142-0858 leo.obrst~boeing.com intelligent agent framework(Decker 1995; Decker, Lesser, Nagendra Prasad, & Wagner 1995; Williamson, Decker & Sycara 1996; Decker & Sycara 1996). Although the project is currently only in an early stage of development, this framework is expected to employ a constraint-centered System Design Management Agent [SDMA]developed by the University of Toronto’s IE Department in conjunction with Boeing (Bilgic et al. 1996; Bilgic, 1997). The SDMA will use the constraint-based Toronto Ontologies for a Virtual Enterprise (TOVE)ontologies (Gruninger & 1995), and its domaintheories for design engineering (Lin, Fox, & Bilgic 1996) and dependent underlying theories, phrased as KIF/Ontolingua assertions (Genesereth &Fikes 1992; Gruber 1993; Gruber 1994, Gruber & Olsen 1994) in an axiomatic system running in the constraint logic system ECLiPSe(Meier 1995), as its primary knowledgeresource to monitor an ongoing design project, offering resourceallocation, coordination-strategic, and risk-abatement suggestions to the interacting agents and humansas best it can. In addition, the underlying domainmodel and agent architecture for MADEsmartwill employ primitive constraint notions to facilitate the design engineering process. These constraints will be evaluated by a given agent at goal commitment time, during planning, scheduling, or execution of tasks. Abstract As part of the DARPA Rapid Design Exploration and Optimization(RaDEO) program,Boeing, Philadelphia, involved in an on-going concurrent design engineering research project called MADEsmart whichseeks to partially automatethe Integrated ProductTeam(IFr) conceptused Boeingfor organizingthe designengineeringprocess, with the aid of intelligent agent technology.Althoughcurrently only in an early stage of development, the project is expectedto crucially employa constraint-centered System DesignManagement Agentdeveloped by the University of Toronto’s IE Departmentin conjunction with Boeing. The SDMA will use the constraint-based TorontoOntologiesfor a Virtual Enterprise (TOVE) ontologies, and its domain theories for design engineeringand dependentunderlying theories, phrased as KIF/Ontolinguaassertions in an axiomatic systemrunning in the constraint logic system ECLiPSe,as its primary knowledgeresource to monitoran ongoingdesign project, offering resource-allocation, coordination-strategic,andrisk-abatementsuggestionsto the interactingagents as best it can. In addition,the underlying domainmodeland agent architecture for MADEsmart will employprimitive constraint notions to facilitate the design engineeringprocess. Theseconstraints will be evaluatedby a given agent at goal commitment time, during planning, scheduling,or executionof tasks. This paper describesthe expected interaction of constraints and agents of this ongoingresearch. Agents, Agent Infrastructure, and Constraints The current modelof an individual agent in MADEsmart is displayed in Figure 1, and is described below. This TAEMS-centricview of the agent considers an intelligent agent as communicating with other agents -- in MADEsmart, using a version of KQML (Finin et al. 1993) -- and being locally responsible for the planning, scheduling, and execution of task structures, though those tasks mayhave non-local effects on other agents or be nonlocally affected by other agents. MADEsmartdiverges somewhatfrom TAEMS in that the system is intended to be an actual, not just simulated, environment for agent interaction with humansand other software systems. A task structure in our view is a hierarchy of subtasks which need to be reduced and eventually instantiated by an Introduction As part of the DARPARapid Design Exploration and Optimization (RaDEO)program, Boeing, Philadelphia, involved in an on-going concurrent design engineering research project called MADEsmart (Barrett et al. 1997; Obrst et al. 1997) which seeks to partially automate the Integrated Product Team(IPT) concept used by Boeing for organizing the design engineering process, with the aid of intelligent agents operating in a blackboard environment (Lander, Staley & CorkiU1996). The core architecture MADEsmart is centered on the Task Analysis, Environment Modeling, and Simulation [TAEMS]-related 83 agent based uponan internalized notion of a goal to be satisfied. Thisis also similarto the unifiedreactive/rational conceptionof agents of (Kowalski&Sadri 1996), whonote correlations to deductivedatabase techniques(Das 1992), and wewill investigate applyingtheir proof procedurefor the agent control cycle (using goal reduction rules and integrity constraints) to the TAEMS model.In our view, an agent receives a KQML request from another agent to achievea particular goal. This interaction is displayedat the top of Figure1, whichalso depicts the requestingagent interacting with events on the blackboard,a representation of the evolving implementationof the agent environment. Therequest also has somecontentassociatedwith it, in the formof constraints andpossiblyother assertions related to the goal, phrased in a K1F-subset(Genesereth & Fikes 1992) language. MADEsmart AgentInternal Figure 1. MADEsmart Agent Internal View The receiving agent first determines whether it can possiblysatisfy the goal -- i.e., does it possessanytask structures (hierarchically composed of subtasksin various ordering relationships, with actions as leaves) - by consulting a rule base of goal-to-task mappingsin its internal knowledgebase, and committingif it determines that it can satisfy the goal request. Oncethe agentcommits to the goal, it continuesby planningwhichtasks to perform, i.e., a reduction andthen instantiation processwhichmay evaluate, with respect to the tasks, certain domain constraints fromthe contentslot of the original request. In Figure1, the constraintflowis representedby a dottedline. In general, wewishto evaluate only domainconstraints in the planning process, and delay the evaluation of performance constraints until the schedulingand execution 84 phase.Becausethe project is still in its infancy,however, weare unsure what constitutes such clearly delineated domainconstraints. It maybe the case, for example,that there are twoalternative task structures to performa given type of analysis, say a stress analysis of a proposeddesign for an aircraft wingpanel, andthe constraint tradeoff is formulatedin terms of granularity of the analysis vs. the time requiredto performthe analysis. Agross analysis will take less time than a finer analysis. It maybe that an agent can makethat decisionearly in the planningprocess, rather than delayit for the schedulingprocess, whereperformance constraintsare evaluated. After planningis completed,the agent locally schedules (in cooperation with a scheduling module which has awarenessof other agents’ tasks) and executes the task structures whichwill satisfy that goal, interacting with other agents during this process if there are non-local effects. Appropriateconstraints maybe evaluated at both scheduling and execution times. As mentionedearlier, these constraints in generalwill be performance constraints and concern notions such as application resources, time (duration, deadline), and quality (a vector of metrics task evaluation). Part of the interaction among agentswill be to coordinate the use of software and hardware applications which facilitate the designandanalysis activities of the design engineeringprocess. Theseinclude both CAD tools, which contain geometricand parametric data on the design, and analysis tools, most of which are legacy applications performingevaluation or optimization of factors such as weight, stress, and producibility. Althoughour initial modelof the type of schedulerwewill employis that most closely correlated to the TAEMS architecture i.e., Designto-Time (Garvey & Lesser 1993) and Design-to-Criteria (Wagner,Lesser, &Garvey1996),wewill also try to link to the constraint-reasoning SDMA, whichwill embodya risk assessor embodiedas a Bayesian network. The SDMA will inspect agents’ rules andtask structure libraries, in addition to communicating directly with the agent, to offer it advice during the planning and scheduling processes. Anotherkind of agent interaction proposedinvolves the SDMA interacting with the agent environmentconcerning moreglobal notions about the ongoingdesignprocess, such as risk-abatement,coordinationstrategies, and resourceallocation schemes, and inspecting the internal data structures of the evolvingdesign network.Finally, other kinds of interaction will facilitate humanand agent collaboration. Figure 2 displays a picture of the proposedMADEsmart agent interaction process. In this schema,each user is associated with a User CoordinationAssistant Agent, an agent whichrepresents the user’s interface to the system, andhis/her interests, views,status, etc., withrespectto ongoing designs and projects. User modelsmayor maynot be globally available (displayed as the upper leftmost knowledge store in Figure2), but will be accessible to the UCAA. The UCAA agents will initially perform the tasks specifiedby the user, but it is anticipatedthat eventually dedicated agents representing functional responsibilities such as stress analysis or organizational entities will be necessary. These are displayed as specialized agents intervening between a given UCAAand individual Custodian/Design Thread agents. These latter agents are specialized information agents who maintain a given project’s evolving design network, an internal representation of connected nodes which contain at least some information on the applicable constraints holding at correlated temporal points in the design evolution and partly representing design alternatives. Another aspect of the interaction between agents and constraints in MADEsmart is in our use of the cost model COSTADE (Mabson et al. 1996), to drive design analysis (including producibility, stress, and weights analysis), Multi-Disciplinary Optimization, and preliminary and detaiIed design. COSTADE is a Fortran-based model we maywish to partially disentangle and map componentsto the constraint logic-based SDMA. contents of this paper remains speculative in nature, hedged with anticipations. The work envisioned is collaborative, and we expect additional crucial contributions from our subcontractors. Although the primary purpose of MADEsmartis to enable agents to facilitate the collaborative design process amonghumans, we have hope that they may fill a larger capacity, as intelligent collaborators themselves working on a design project represented as a networkof constrained design alternatives. Acknowledgements This work is supported by cooperative agreement 70NANB6H0074 under the DARPARaDEOprogram, with the majority of the funding from DARPA(with NIST participating and acting as their agent) and the remaining funding supplied by Boeing. References Barrett, Tom;Coen, Gary; Hirsh, Joel; Obrst, Leo; Spering, Judith; Trainer, Asa. 1997. MADEsmart:An Integrated Design Environment. Submitted to 1997 ASMEDesign for Manufacturing Symposium. Bilgic, Tarter. 1997. Product Data ManagementSystems: State-of-the-Art and the Future. Enterprise Integration Laboratory, Department of Mechanical and Industrial Engineering, University of Toronto, manuscript, January 28, 1997. Bilgic, T; Chionglo, J.; Fox, M.; Gupta, L; Gwidzka, J.; Leizorewiez, W.; Lin, J. 1996 Knowledge-AidedDesign for Design-in-the-large. Enterprise Integration Laboratory, Department of Mechanical and Industrial Engineering, University of Toronto. Technical Report, June, 1966. Das, Subrata Kumar. 1992. Deductive Databases and Logic Programming. Workingham, England and Reading, MA:Addison-Wesley. Decker, Keith. 1995. Environment Centered Analysis and Design of Coordination Mechanisms.Ph.D diss., Dept. of ComputerScience, University of Massachusetts. Decker, Keith; Lesser, V.R.; Nagendra Prasad, M.V.; Wagner, T. 1995. MACRON: An Architecture for MultiAgent Cooperative Information Gathering. In Proceedings of the International Conference on Information and Knowledge Management (CIKM-95) Workshop Intelligent Information Agents, Baltimore, MD. Figure 2. MADEsmart Agent External View Conclusion Decker, Keith; Sycara, Katia. 1996. Designing Reusable Behaviors for Information Agents. Manuscript, The Robotics Institute, Carnegie-MellonUniversity. MADEsmart is very mucha research project still evolving. Weare just out of the first six-month period and expect to witness much more change over the course of this prospectively three-year project. Hence, much of the 85 Finin, T.; Weber, J; Widerhold, G.; Genesereth, M; Fritzson, R.; McKay,D.; Mcquire, J.; Pelavin, R.; Shapiro, S.; Beck, C. 1993. Specification of the KQML Agent Communication Language. Draft, June 15, 1993. The DARPA KnowledgeInitiative External Interfaces Working Group. Z.B. 1996. Cost Optimization Software for Transport Aircraft Design Evaluation (COSTADE). NASA Contractor Report 4736. Garvey, Alan and Lesser, Victor. 1993. Design-to-Time Real-Time Scheduling. IEE Transactions on Systems, Man and Cyhemettics, 23(6): 1491-1502. Obrst, Leo; Woytowitz, Michael; Rock, Dennis; Lander, Susan; Gallagher, Kevin; Decker, Keith. 1997. Agentbased Integrated Product Teams. Submitted to 1997 ASME Design Engineering and Computers in Engineering Conference, Engineering Information Management Symposium. Meier, Micha. 1995. ECLiPSe 3.5 User Manual. ECRC, Munich, Germany, December, 1995. Genesereth, Michael R.; Fikes, Richard. 1992. Knowledge Interchange Format Version 3.0 Reference Manual. Logic Group, Dept. of ComputerScience Stanford University. Wagner, Thomas; Garvey, Alan; and Lesser, Victor. 1996. Satisficing Evaluation Functions: The Heart of the New Design-to-Criteria Paradigm. In UMASS Department of Computer Science Technical Report TR-1996-82, November, 1996. Gmber, Tom. 1993. A Translation Approach to Portable OntologySpecifications. KnowledgeAcquisition, V. 5, pp. 199-220. Gruber, Tom. 1994. TowardPrinciples for the Design of Ontologies Used for KnowledgeSharing. In Guarino, N., and Poli, R., eds. FormalOntology in Conceptual Analysis and KnowledgeRepresentation. Williamson, Mike; Decker, Keith; Syeara, Katia. 1996. Unified Information and Control Flow in Hierarchical Task Networks. Technical report, The Robotics Institute, Carnegie-Mellon University. Gruber, Tom; Olsen, Gregory R. 1994. An Ontology for Engineering Mathematics. In Jon Doyle, Piero Torasso, and Erik Sandewall eds. Fourth International Conference on Principles of KnowledgeRepresentation and Reasoning, Gustav Stresemann Institut, Bonn, Germany, Morgan Kaufman. Available as Stanford Knowledge Systems Laboratory technical report KSL-94-18. Gruninger, M; Fox, M. 1995. Methodologyfor the Design and Evaluation of Ontologies. Department of Industrial Engineering, University of Toronto. Kowalski, Robert; Sadri, Fariba. 1996. Towardsa Unified Agent Architecture that Combines Rationality with Reactivity. Department of Computing, Imperial College, manuscript, June 12, 1996. Lander, Susan, E; Staley,Scott M.; Corkill, Daniel D. 1996 Designing Integrated Engineering Environments: Blackboard-based Integration of Design and Analysis Tools. Concurrent Enginnering: Research and Applications, Special Issue on the Application of Multiagent Systems to Concurrent Engineering, 4(I):59-72, March, 1996. Lin, Jinxin; Fox, Mark S.; Bilgic, Taner. 1996. A Requirement Ontology for Engineering Design. Enterprise Integration Laboratory, Department of Mechanical and Industrial Engineering, University of Toronto, manuscript, September, 1996. Mabson, G.E.; Ilcewicz, L.B.; Graesser, D.L.; Metschan, S.L.; Proctor,M.R.; Tervo, D.K.; Turtle, M.E.; Zabinsky, 86