DESIGNING AN EXTENDED SET OF COORDINATION MECHANISMS FOR MULTI-AGENT SYSTEMS by Wei Chen A dissertation submitted to the Faculty of the University of Delaware in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Computer Science Fall 2005 c 2005 Wei Chen All Rights Reserved UMI Number: 3200551 Copyright 2005 by Chen, Wei All rights reserved UMI Microform 3200551 Copyright 2006 by ProQuest Information and Learning Company All rights reserved This microform edition is protected against unauthorized copying under Title 17, United States Code ProQuest Information and Learning Company 300 North Zeeb Road P.O Box 1346 Ann Arbor, MI 48106-1346 DESIGNING AN EXTENDED SET OF COORDINATION MECHANISMS FOR MULTI-AGENT SYSTEMS by Wei Chen Approved: B David Saunders, Ph.D Chairperson of the Department of Computer and Information Sciences Approved: Thomas M Apple, Ph.D Dean of the College of Arts and Sciences Approved: Conrado M Gempesaw II, Ph.D Vice Provost for Academic and International Programs I certify that I have read this dissertation and that in my opinion it meets the academic and professional standard required by the University as a dissertation for the degree of Doctor of Philosophy Signed: Keith S Decker, Ph.D Professor in charge of dissertation I certify that I have read this dissertation and that in my opinion it meets the academic and professional standard required by the University as a dissertation for the degree of Doctor of Philosophy Signed: Daniel L Chester, Ph.D Member of dissertation committee I certify that I have read this dissertation and that in my opinion it meets the academic and professional standard required by the University as a dissertation for the degree of Doctor of Philosophy Signed: Chien-Chung Shen, Ph.D Member of dissertation committee I certify that I have read this dissertation and that in my opinion it meets the academic and professional standard required by the University as a dissertation for the degree of Doctor of Philosophy Signed: Matthew J Hoffmann, Ph.D Member of dissertation committee ACKNOWLEDGMENTS I would like to thank all my colleagues in our multi-agent system group of the Department of Computer and Information Sciences, University of Delaware, especially the leader and my adviser, Dr Keith S Decker My deepest thanks to Keith for his guidance and support, especially for his patience and help throughout the process of my dissertation Without these, my research objectives would never have succeeded quickly and well His outstanding ideas in this research area have given rise to various projects, including this dissertation I would like to give very special thanks to Dr Daniel L Chester, who pointed out most of the mistakes in early versions of this dissertation and provided many constructive ideas; his expertise and interest in various research fields is a great inspiration for my life and for my future research work Deep thanks to Dr Chien-chung Shen for his sharp questions and directions; I learned a lot from his character and serious attitude towards research Many thanks to Dr Matthew J Huffmann for his guidance and insights from a special perspective, which helped open up this dissertation to readers outside of distributed artificial intelligence I also offer special thanks to the researchers in the MAS group of the University of Massachusetts at Amherst I went to the UMass campus for ideas during the early stage of my research work and got in touch with some of their thoughts My advisor’s advisor, Dr Victor Lesser, showed me the potential development in the field of multi-agent systems and encouraged me to continue work in this research area iv I should present special thanks to my wife, Guangmeng (Grace), for her continuous support and all the love, care, and help in my life I am profoundly thankful to my parents; I hope they will remain happy, healthy, and proud of their son v TABLE OF CONTENTS LIST OF FIGURES xii LIST OF TABLES xvi ABSTRACT xvii Chapter INTRODUCTION: WHAT AND WHY? 1.1 1.2 1.3 1.4 What Is Coordination? Why study coordination? Coordination in Multi-Agent Systems Coordination in a Domain Application—Emergency Medical Service Systems 1.4.1 1.5 14 Three Coordination Points in EMS 16 Intellectual Contributions 19 1.5.1 1.5.2 1.5.3 1.5.4 1.5.5 1.6 Problem Scope Formal Representation An Extended Set of GPGP Coordination Mechanisms Architectural Support Experimentation and Qualitative Analysis 21 23 24 27 28 Chapter Summary 30 vi RELATED WORK 31 2.1 Coordination Approaches in Multi-Agent Systems 31 2.1.1 Coordination Science Presented From MIT 33 2.1.1.1 2.1.1.2 2.1.2 Direct Inspiring Base — TÆMS and GPGP 37 2.1.2.1 2.1.2.2 2.1.2.3 2.1.3 Ideas and Models 46 Coordination Selection Strategy 47 Teamwork 48 2.1.4.1 2.1.4.2 2.2 2.3 TÆMS 37 GPGP 39 Other work related to TÆMS and GPGP 42 Jennings and Colleagues 45 2.1.3.1 2.1.3.2 2.1.4 Key Ideas and Coordination Methods 33 Applications 37 Teamwork Model and Agent Architecture 50 Current Work 51 Other Approaches 52 Summary 59 DEVELOPING AN EXTENDED SET OF GPGP COORDINATION MECHANISMS 61 3.1 3.2 Extending Traditional GPGP Coordination Approach 61 Features of Our Extended GPGP Approach 63 3.2.1 3.2.2 Task Structures and Enables Relationship 63 Real time task structure alteration 66 vii 3.2.3 3.3 Abstraction Level 69 An Extended Set of GPGP Coordination Mechanisms for Enables Relationships 70 3.3.1 3.3.2 3.3.3 3.3.4 3.3.5 3.3.6 3.3.7 3.3.8 3.3.9 3.3.10 3.3.11 3.3.12 3.3.13 3.3.14 3.3.15 3.3.16 3.3.17 Avoidable Dependency Sacrifice Avoidable Dependency Coordination by Reservation Predecessor Earliest Start Time (EST) Commitment Predecessor Deadline Commitment Predecessor Notice at Start Predecessor Sending Result Successor Deadline Commitment Successor Earliest Start Time (EST) Commitment Demotion Shift Promotion Shift Polling for Result Polling for Schedule Constant Headway / Timetabling Third Party Execution Third Party Coordinator Bidding 73 76 77 78 80 81 82 82 84 84 86 86 87 88 89 90 91 FORMAL REPRESENTATION AND AGENT ARCHITECTURAL SUPPORT 93 4.1 Extended Hierarchical Task Networks—EHTNs 93 4.1.1 4.1.2 4.2 4.3 4.4 4.5 Introduction 93 Extending HTNs (EHTN) to Represent Coordination Problems 96 Agent Architectural Support and A Schedule Coordination Problem GPGP Mechanisms Coordination Mechanism Selection The Recast of Selected GPGP Coordination Mechanisms Using EHTNs 4.5.1 103 108 112 117 Definitions 120 viii 4.5.2 Avoidable Dependency 121 4.5.2.1 4.5.3 Coordination by Reservation 4.5.3.1 4.5.3.2 4.5.4 4.5.5 4.5.6 4.6 4.7 Re-writing of the Extended HTNs 123 123 Re-writing of the Extended HTNs 125 Coordination Protocol 127 Demotion Shift Dependency 129 Coordination by Sending Result 130 Coordination by Polling Result 133 Implementation 133 Summary about Formal Representation 135 EVALUATION 137 5.1 5.2 5.3 5.4 5.5 Modeling Task Environment General Approach Adjust Coordination Behaviors Under Changing Environmental Factors Coordination as an Independent Component Analysis of Coordination in Emergency Medical Services 5.5.1 5.5.2 5.5.3 5.5.4 138 139 142 142 144 144 145 147 150 Framework Overview Introduction to Participating Agents EMS System Operation Overview System Development The Application of Coordination Mechanisms to Three Major Coordination Points in EMS Random Distribution for Incident Generation, Agent Working Time, and More 150 153 157 159 Background Introduction of Emergency Medical Service (EMS) EMS Model Emergency Medical Service (EMS) Framework 5.5.4.1 5.5.4.2 5.5.4.3 5.5.4.4 5.5.4.5 5.5.4.6 ix 161 163 [75] Babara Grosz and Sarit Kraus Collaborative plans for complex group action Artificial Intelligence, 86(2):269–357, 1996 [76] B.J Grosz and C.L Sidner Plans for discourse In P.R Cohen, J Morgan, and M.E Pollack, editors, Intentions in Communication MIT Press, 1990 [77] G.Weiss Multiagent Systems, A Modern Approach to DAI, pages 151–157 MIT Press, 1999 [78] G.Weiss Multiagent Systems, A Modern Approach to Distributed Artificial Intelligence, pages 88–92, 589 MIT Press, 1999 [79] B Hayes-Roth, L Brownston, and R.V Gen Multiagent collaboration in directed improvisation In Victor Lesser, editor, ICMAS95, pages 148–154, 1995 [80] A Herzig and D Longin A logic of intention with cooperation principles and with assertive speech acts as communication primitives In Proceedings of the Second AAMAS, 2002 [81] Carl Hewitt Offices are open systems ACM Transactions on Office Information Systems, 4(3):271–287, July 1986 [82] Eric J Horvitz Reasoning under varying and uncertain resource constraints In Proceedings of the Seventh National Conference on Artificial Intelligence, August 1988 [83] B A Huberman Open systems: The ecology of computation PARC working paper, page 74, 1988 Palo Alto, CA: Xerox PARC [84] B A Huberman The Ecology of Computation 1988 Amsterdam: North-Holland [85] Wafik H Iskander Simulation modeling for emergency medical service systems In Proceedings of the 1989 Winter Simulation Conference, 1989 [86] M Miller et al J Yen, T Ioerger Cast: Collaborative agents for simulating teamwork 2001 [87] N R Jennings Commitments and conventions: The foundation of coordination in multi-agent systems The Knowledge Engineering Review, 8(3):223–250, 1993 [88] N R Jennings Coordination techniques for distributed artificial intelligence In Foundations of Distributed Artificial Intelligence, pages 187–210 Wiley, 1996 [89] Nicholas R Jennings Controlling cooperative problem solving in industrial multiagent systems using joint intentions Artificial Intelligence, 75(2):195–240, 1995 208 [90] H Kitano and et al Robocup: The robot world cup initiative In IJCAI05 Workshop on Entertainment and AI life, 1995 [91] V Lesser, K Decker, N Carver, A Garvey, D Neimen, M.V Prassad, and T Wagner Evoolution of the gpgp domain independent coordination framework Technical Report 1998-005, CS UMASS, January 1998 [92] V Lesser, B Horling, R Vincent, A Raja, and S Zhang The taems white paper http://mas.cs.umass.edu/research/taems/white, 2005 [93] Victor Lesser Reflections on the nature of multi-agent coordination and its implications for an agent architecture Autonomous Agents and Multi-Agent Systems, 1:89–111, 1998 [94] Victor R Lesser Coordination in distributed problem solving networks DARPA proposal, COINS, University of Massachusetts, 1989 [95] H Li, E Durfee, and K Shin Multiagent planning with internal execution resource constraints In In Proceedings of the Second AAMAS, July 2003 [96] K Lodding and P Brewster Multi-agent organisms for persistent computing In In Proceedings of the Third AAMAS, 2004 [97] R Maheswaran, P Pearce, P Varakantham, E Bowring, and M Tambe Valuations of possible states (vps): A quantitative framework for analysis of privacy loss among collaborative personal assistant agents In AAAI Spring Symposium on Persistent Assistants: Loving and Working with AI, Menlo Park, CA, 2005 [98] Thomas Malone and Kevin Crowston The interdisciplinary study of coordination In ACM Computing Surveys, pages 87–119, March 1994 [99] Thomas W Malone Modeling coordination in organizations and markets Management Science, 33:1317–1332, 1987 [100] Thomas W Malone What is coordination theory? In Proceedings of the National Science Foundation Coordination Theory Workshop, February 1988 [101] Thomas W Malone, Kevin Crowston, Jintae Lee, and Brian Pentland Tools for inventing organizations: Toward a handbook of organizational processes Center for Coordination Science Technical Report 141, MIT Sloan School of Management, 1993 [102] M Mamei and F Zambonelli Field-based motion coordination in quake arena In In Proceedings of the Third AAMAS, 2004 209 [103] C.E Martin Adaptive Decision-Making Frameworks for Multiagent Systems PhD thesis, Electrical and Computer Engineering, University of Texas at Austin, Austin, Texas, 2001 [104] Bureau of Labor Statistics Occupational outlook handbook, 2002-2003 edition, 2003 http://www.bls.gov/oco/ [105] H Nagatuma Developing of an emergency medical video multiplexing transport system Journal of Medical Systems, 27(2):133–140, 2003 [106] L Ngo, P Haddawy, and H Nguyen A modular structured approach to conditional decision-theoretic planning In AIPS98, pages 111–118, 1998 [107] Tim Oates, M Schmill, and Paul Cohen Identifying qualitatively different outcomes of actions: Gaining autonomy through learning In Proceedings of Autonomous Agent 2000, pages 110–111, 2000 [108] W Parker and R Johnson Simulation of a coordinated accident rescue system In the fourth annual conference on Applications of simulation, 1970 [109] R Parr Flexible decomposition algorithms for weakly coupled markov decision problems In Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence (UAI-98), Morgan Kaufmann, 1998 [110] B Pimentel, G Pereira, and M Campos On the development of cooperative behavior-based mobile manipulators In Proceedings of the Second AAMAS, 2002 [111] M.V Nagendra Prasad and V.R Lesser Learning situation-specific coordination in generalized partial global planning In AAAI Spring Symposium on Adaptation, Co-evolution and Learning in Multiagent Systems, Stanford, March 1996 [112] D Pynadath and M Tambe An automated teamwork infrastructure for heterogeneous software agents and humans Autonomous Agents and Multiagent Systems, 2002 [113] D Pynadath and M Tambe Multiagent teamwork: Analyzing the optimality and complexity of key theories and models In The First International Conference on Autonomous Agents and Multiagent Systems, pages 873–880, 2002 [114] D Pynadath and M Tambe Team coordination among distributed agents: Analyzing key teamwork theories and models In Proceedings of the AAAI Spring Symposium on Intelligent Distributed and Embedded Systems, 2002 210 [115] M Tambe R Maheswaran and E Bowring Taking dcop to the real world: Efficient complete solutions for distributed multi-event scheduling In Proceedings of AAMAS 2004, pages 310–317, 2004 [116] A.S Rao, A Lucas, D Morley, M Selvestrel, and G Murray Agent-oriented architecture for aircombat simulation Technical Report 42, iAustralian Artificial Intelligence Institute, 1993 [117] Charles Rich and Candace L Sidner COLLAGEN: When agents collaborate with people In W Lewis Johnson and Barbara Hayes-Roth, editors, Proceedings of the First International Conference on Autonomous Agents (Agents’97), pages 284– 291, New York, 5–8, 1997 ACM Press [118] S Russell and P Norvig Artificial Intelligence a Modern Approach, Second Edition Prentice Hall, 2002 [119] Stuart Russell and Eric Wefald Do the Right Thing: Studies in Limited Rationality MIT Press, Cambridge, MA, 1991 [120] S Rustogi and M Singh Be patient and tolerate imprecision: How autonomous agents can coordinate effectively In Proceedings of IJCAI’99, Stockholm, Sweden, August 1999 [121] Gordon M Sachs Officer’s Guide to Fire Service EMS Fire Engineering, 1999 [122] S Sen, S Roychoudhury, and N Arora Effect of local information on group behavior In Proceedings of the International Conference on MAS, pages 315–321, 1996 [123] B Singh Interconnected roles (ir): A coordination model Technical Report CT084-92, Microelectronics and Computer Technology Corp (MCC)., Austin, TX, 1992 [124] Munindar P Singh A customizable coordination service for autonomous agents In Proceedings of ATAL97, 1997 [125] Munindar P Singh Interaction-oriented programming for the web In Cooperative Information Systems (CoopIS), May 1997 [126] Munindar P Singh Developing formal specifications to coordinate heterogeneous autonomous agents In ICMAS98, pages 261–268, 1998 [127] Munindar P Singh Synthesizing coordination requirements for heterogeneous autonomous agents Autonomous Agents and Multi-Agent Systems, 3(2):107–132, June 2000 211 [128] E Sklar, M Davies, and M Co Simed: Simulating education as a multi-agent system In In Proceedings of the Third AAMAS, New York, USA, 2004 [129] I.A Smith and P.R Cohen Toward a semantics for an agent communication language based on speech acts In Proceedings of the Thirteenth National Conference on Artificial Intelligence, pages 24–31, August 1996 [130] Y So and E Durfee Designing organizations for computational agent In K.M Carley In M.J Pritula and L Gasser, editors, Simulating Organizations, pages 47–64 AAAI press, 1998 [131] Antonella Stefano and Corrado Santoro The coordination infrastructure of the arca framework In Proceddings of Autonomous Agent 2000, pages 173–174, 2000 [132] M Tambe Recursive agent and agent group tracking in a real-time, dynamic environment In Proceedings of the First International Conference on Multi-Agent Systems, pages 368–375, San Francisco, June 1995 AAAI Press [133] M Tambe Agent architectures for flexible, practical teamwork In Proceedings of the Fourteenth National Conference on Artificial Intelligence, Providence, July 1997 [134] M Tambe Teamwork in real-world, dynamic environments In Proceedings of the International conference on multi-agent systems(ICMAS96), pages 361–368, 1997 [135] M Tambe Towards flexible teamwork Journal of Artificial Intelligence Research, 7:83–124, 1997 [136] M Tambe, W Johnson, R Jones, F Koss, J Laird, P Rosenbloom, and K Schwamb Intelligent agents for interactive simulation environments AI Magazine, 16(1):15–39, 1995 [137] M Tambe, G Kaminka, S Marsella, I Muslea, and T Raines Two fielded teams and two experts: A robocup response challenge from the trenches In IJCAI99, 1999 [138] Milind Tambe Eliminating Combinatorics from Production Match PhD thesis, School of Computer Science, Carnegie Mellon University, 1991 Available as technical report CMU-CS-91-150 [139] Milind Tambe and Paul Rosenbloom A framework for investigating production system formulations with polynomially bounded match page 693 [140] T.Hogg and B.Huberman Controlling chaos in distributed systems In IEEE Transactions on Systems, Man, and Cybernetics, pages 1325–1332, 1991 212 [141] T Wagner V Lesser, K Decker Evolution of the gpgp/t?ms domain-independent coordination framework Autonomous Agents and Multi-Agent Systems, 2004 [142] T Wagner and et al Coordinators: Coordination managers for first responders In Proceedings of the Third AAMAS03, New York, USA, 2004 [143] T Wagner, A Garvey, and V Lesser Complex goal criteria and its application in design-to-criteria scheduling In Proceedings of the Fourteenth National Conference on Artificial Intelligence, pages 294–301, Providence, 1997 [144] Thomas Wagner and Victor Lesser Toward generalized organizationally contexted agent control Technical Report TR99-18, UMASS CS, April 1999 [145] R Wears and C Winton Simulation modeling of prehospital trauma care In the 1993 Winter Simulation Conference, 1993 [146] Michael Wellman A market-oriented programming environment and its application to distributed multicommodity flow problems Journal of Artificial Intelligence Research, 1:1–23, 1993 [147] M Williamson, K S Decker, and K Sycara Unified information and control flow in hierarchical task networks In Proceedings of the AAAI-96 workshop on Theories of Planning, Action, and Control, 1996 [148] S Willmott and B Faltings The benefits of environment adaptive organizations for agent coordination and network routing problems In Proceedings of IJCAI’99, Stockholm, Sweden, August 1999 [149] R Wulfhorst, L Nakayama, and R Vicari A multiagent approach for musical interactive systems In AAMAS, pages 584–591, 2003 [150] George Wyner and Jintae Lee Defining specialization for process models Technical Report 216, 2001 [151] P Xuan and V Lesser Multi-agent policies: from centralized ones to decentralized ones In AAMAS02, pages 1098–1105 [152] Q Yang Formalizing planning knowledge for hierarchical planning Computational Intelligence, 6(1):12–24, February 1990 [153] R Young, M Pollack, and J Moore Decomposition and causality in partial-order planning In Proc 2nd Intl Conf on A.I Planning Systems, pages 188–193, June 1994 213 Appendix A PROOF OF EXPRESSIVENESS THEOREM This appendix contains the proof of the Expressiveness Theorem The definition of expressiveness for knowledge representation languages is presented in [5] and the adaptation of this definition to planning languages is introduced in [60] The notion of expressiveness is based on the idea that if a language L1 can be expressed by another language L2 , then for any set of sentences Γ1 in L1 , there must be a corresponding set of sentences Γ2 in L2 with the same meaning In [5], Baader captures “the same meaning” for knowledge representation languages by requiring that the two sets of sentences have the same set of models Based on the definition of equivalence of HTN models in [60], it is easy to conclude that when a planning language L1 can be expressed by a planning language L2 , but not vice versa, then L2 is strictly more expressive than L1 Suppose that ω is a symbol translation function, ψ is a sentence translation function, L is a traditional HTN planning language A semantic structure for a traditional HTN has the same form as a semantic structure for our EHTN, with some restrictions The semantic structure of the traditional HTN is a triple M = SM , FM , TM , where S is the set of states, F interprets actions as state transitions, and T interprets non-primitive tasks as sets of ground primitive task networks, with the following restrictions: Since the traditional HTN lacks the notion of information flow, control flow, and quantitative representation, non-primitive tasks consist of only goal tasks, and T is not defined for information flow, control flow, and quantitative representation, etc 214 Given the above definitions (further explanation about model-theoretic expressiveness is in [5] and [60]), we prove the expressiveness theorem by stating two Lemmas first In order to show that our EHTN is more expressive than the traditional HTN, we need to show that the traditional HTN planning language can be expressed by our EHTN planning language, while our EHTN can not be expressed by the traditional HTN The EHTN is based on the traditional HTN Thus, the EHTN includes all the planning language constructs of the traditional HTN On the other hand, the extended HTN contains extra models for information flow, control flow, and quantitative representation As a result, our EHTN does not require any extra symbols for representing the traditional HTN Thus, EHTN will use a superset of symbols S, which contains two parts: (1) exactly the same set of constants, predicates, and actions (primitive tasks) as the traditional HTN; (2) symbols for the extra part that traditional HTN does not have, e.g., non-local task NLT , provision I, outcome O, task and action attributes A, and task and action characteristic accumulation functions, which provides additional representation for information flow, control flow, quantitative representation Given a set of the traditional HTN operators Γ, we define ψ(Γ): (1) ψ(Γ) contains a super set of operators as Γ (2) For each goal task achieve[l, φ] and each action f , ψ(Γ) contains the method (achieve[l, φ], [(n1 : do[f, φ])(n2 : achieve[l, φ]), (n1 ≺ n2 ), φ, Lφ , Aφ ]) Notably, (2) contains the new language model for the EHTN The proof of this Lemma is similar to that described in [60] There is an implicit method for each goal task stating it can be expanded to a dummy task when the goal literal is true, with the methods in ψ(Γ), any goal task can be expanded to any sequence of actions, which would be a plan for the goal task whenever all the actions are executable and the goal literal is true in the end of the plan Here we explicitly point out that the methods reflect the restrictions on the models 215 of the traditional HTN planning is a subset of the models of EHTN because flow representation and quantitative analysis of tasks and actions is not available; as a result Γ and ψ(Γ) have the same set of models except that there are empty models (discarding the quantitative representation, information flow, and control flow) in the superset for the EHTN Thus, we conclude Lemma The traditional HTN language can be expressed by the extended HTN with respect to model theoretic expressivity The converse of this Lemma is not true The proof is exactly like the proof of HTN cannot be expressed by STRIPS in [60] with the difference as follows Here, the lack of a computable transformation from our EHTN to the traditional HTN is a key point Just like the analogy that associates STRIPS-style grammars with regular grammars, and the traditional HTN with context-free grammars, our extended HTN is like attribute grammars1, which are actually context-free grammars extended with additional two features—one for data and one for computation The attributes are used to hold information of the semantic content attached to the specific node in an Abstract Syntax Tree (AST) All nodes in the AST that correspond to the same grammar symbol S have the same formal attributes, but their values, the actual attributes, may differ At the same time, characteristic accumulation functions stand for the computation of an attribute grammar Given an extended HTN domain Γ that corresponds to a attribute grammar but not context-free grammar, Γ will have a minimum model2 M such that the interpretation TM from non-primitive tasks to sets of primitive tasks and non-local tasks will map compound task symbols to sets of totally ordered primitive task networks with attributes associated with every node in the task networks Since TM maps compound tasks into an attribute grammar but not just into a context-free grammar, no traditional HTN planning domain can have a model equivalent to M Thus we state the following Lemma: Formal proof of the relationship between EHTNs and attribute grammars is out of the scope of this dissertation Refer to the proof of Theorem in [60] 216 Lemma The extended HTN language cannot be expressed by the traditional HTN language with respect to model theoretic expressivity Based on the above two Lemmas, we conclude the following theorem: Expressiveness Theorem The EHTN language L is strictly more expressive than the traditional HTNs 217 Appendix B TASK STRUCTURE EQUIVALENCE AND REDUCTION This appendix discusses task structure reduction Task reduction means that nonstandard-designed task structures can be automatically altered into different forms of task structures that our coordination mechanisms understand The altered task structures are equivalent to the original structures Assumption The general structure of a real Non-Local Task (NLT) that represents a coordination point is as shown in Figure B: There are other task structures that contain Task OK IN (Q,C,D) : Behavior Profile AND AND/OR : CAF OK Ask IN OK IN (Qa21,Ca21,Da21) Handle OK (Qa22,Ca22,Da22) : Task : Action 111 000 000: 111 NLT IN : Provision Cells 111 000 000OK 111 Non−Local Task (Q?, C?, D?) Figure B.1: A real structure that represents a coordination point Non-Local Tasks (NLTs) However, those NLTs are for communication purposes only and there is no dependency associated with the structures, such as in Figure B We claim that most of the structures that contain NLTs can be classified into two groups: coordination points as in Figure B and non-coordination points as in Figure B Currently, we assume that these coordination points were independent Otherwise, if some of the coordination points are dependent/related to each other, they could be 218 Task IN (Q,C,D) : Behavior Profile OK AND/OR : CAF AND OK : Provision Cells : Task Ask IN NLT OK (Qa21,Ca21,Da21) IN 111OK 000 000 111 (Q?, C?, D?) : Action 111: Non−Local Task 000 111 000 Figure B.2: A task structure whose only purpose is to transmit messages; therefore, it is not a coordination point merged into a single NLT, given that DECAF is capable of delivering multiple messages from a single NLT [73] Further in this discussion, we will show that some of the task structures are equivalent and some other task structures can be re-written so that our extended set of GPGP coordination understands them These different appearing, but actually equivalent, task structures occur in the planning time of the agents’ activities by different programmer A programmer who is not familiar with DECAF system may likely plan the tasks for their agents in a “non-standard way” For example, in Figure B, structure (A) is equivalent to structure (B) Based on this equivalence assumption, some task structures can be reduced to the forms that our extended set of GPGP coordination mechanisms understand, so that coordination process can be applied on these structures One example of reduction is shown in Figure B 219 TaskA TaskA OK OK OR OR A1 Sub2 OK OK (Qa1,Ca1,Da1) A1 Equivalent AND Ask OK (Qa1,Ca1,Da1) Ask OK Handle OK IN (Qa21,Ca21,Da21) (Qa21,Ca21,Da21) (Q,C,D) : Behavior Profile IN AND/OR : CAF (Qa22,Ca22,Da22) IN 111OK 000 111 000 Handle OK (Qa22,Ca22,Da22) NLT 000 111 111OK 000 000 111 (Q?, C?, D?) OK : Provision Cells NLT IN OK : Task : Action (Q?, C?, D?) 000: 111 111 000 Non−Local Task Figure B.3: Equivalent task structures while different in the planning time Task1 OK IN AND IN SubTask1 OK IN (Qa21,Ca21,Da21) SubTask2 OK (Qa22,Ca22,Da22) NLT IN 111 000 000OK 111 (Q?, C?, D?) IN SubTask1 OK IN (Qa21,Ca21,Da21) SubTask1 OK (Qa21,Ca21,Da21) NLT Reduction IN (Q,C,D) : Behavior Profile 111 000 111 000OK AND/OR : CAF (Q?, C?, D?) OK OK IN : Action 111 000 111: 000 AND IN SubTask1 OK IN (Qa21,Ca21,Da21) SubTask1 OK IN (Qa21,Ca21,Da21) 111 000 000OK 111 (Q?, C?, D?) Non−Local Task SubTask1 OK (Qa21,Ca21,Da21) NLT IN : Provision Cells : Task Task1 NLT IN 111 000 000OK 111 (Q?, C?, D?) Figure B.4: Reduction of certain task structures into the coordination understandable form 220 Appendix C DISCUSSION OF INCIDENT GENERATION AND SELECTION In Section 5.5.5.3, we briefly mentioned that incident generation and selection is based on a pair-wise comparison It means that during the generation of a series of randomly generated incidents following an exponential distribution, the corresponding seeds for the exponential distribution are recorded and used for future incident generation Thus, when applying a certain coordination mechanism to a particular coordination point, the system performance with the application of the mechanism can be regarded as a direct comparison with that of base case where no any GPGP coordination mechanism is used The major reason is that if the experiments are based on different set of randomly generated incidents (following an exponential distribution), the experimental results can not inform us about the significant difference between applying and not applying a coordination mechanism In addition, it introduces difficulty in analyzing the experimental results That is to say, if the experiments are based on different incidents generated on the fly, we can use the average result values to compare the performances of different mechanisms However, the standard deviation for the average value based on the experimental results from one run of experiment is not satisfactory For example, for a single run of an early experiment on bidding (not shown in this dissertation) the average of a series of response time is 2332 milliseconds based on a set of one hundred randomly generated incidents following an exponential distribution (the mean is 10 seconds); however, the standard deviation of this set of results, the response times, is 1610 milliseconds, which is not a satisfactory at all This extended standard deviation is caused by the geographical 221 features of the local area For example, if an incident is generated at a certain location, and it happens that a police agent arrives at that location when patrolling, thus, the response time for this case is simply zero; in another example, if an incident happens at location twenty-five (index 25) in the local map in Figure 5.2; and the other two agents happen to be far away from that location, it is easy to see that it will cost much more time for any police agent to reach that location Thus, the geographical feature simply hinds the feature of selecting a suitable coordination mechanism The goal of this dissertation is to demonstrate the effectiveness of our novel set of GPGP coordination mechanisms and the performances of these mechanisms in the EMS framework Thus, pair-wise comparison by keeping a series of seeds for an exponential distribution is used as an experiment setting Additionally, pair-wise comparison is employed on all randomly generated factors introduced in Section 5.5.5.3 following corresponding exponential distributions 222 [...]... presents an expressive formalism, projects an extended set of coordination mechanisms, and finally explores the autonomy of coordination behaviors 1.3 Coordination in Multi- Agent Systems My research is focused on the aforementioned coordination problem in complicated multi- agent systems We project the rules and protocols to decide how agents interact with each other using various coordination mechanisms. .. over multiple agents by annotating tasks and actions quantitatively We design and implement an extended set of GPGP (Generalized Partial Global Planning) coordination mechanisms, which are recast using this formalism Second, mechanism development: constructing a large number of mechanisms to deal with the dependencies between multiple agents’ tasks We have catalogued seventeen GPGP coordination mechanisms. .. complicated multi- agent systems Within such systems, if one agent s action(s) dependents on another agent s action(s), we call this kind of relationship a coordination relationship Further, we name the inter -agent relationship inter-dependency2 defined as a relationship between a local task (of one agent) and a non-local task (of another agent) where the execution of one changes some performancerelated... Coordination Mechanisms on EMS Coordination Point Two 183 5.14 Average coordination cost of selected GPGP Coordination Mechanisms on EMS Coordination Point Two 184 5.15 Response time of GPGP Coordination Mechanisms on EMS Coordination Point Three 185 5.16 Average survival rate of selected GPGP Coordination Mechanisms on EMS Coordination. .. spent for the plastic slides, instead of expense or prices, etc 5 The assistant is capable of more tasks, of course, but none of those is our interest for this example 10 Given the fact that the professor’s task, PrintSlides, can not be started until the assistant’s task, FetchSlidesFromDepartmentOffice, is finished, we call the professor an enablee agent and the assistant an enabler agent, meaning... discussing an interdependency relationship, a predecessor is regarded as an enabler agent, and a successor is regarded as an enablee agent Reservation scheme is another approach that can be used to handle the uncertainty problem For example, the professor informs the assistant about the task of bringing the slides and reserves a certain period of the assistant’s time, say, between 1:00 PM and 1:15 PM for. .. local ambulance agent about the victim and the police agent needs to wait for the arrival of an ambulance In the meantime the police agent can not start to carry out other tasks, e.g., to respond to another incident, until the arrival of the ambulance at the incident location It is easy to figure out that the dependency is between the police agent s tasks and the arrival of the ambulance for immediate... rate of selected GPGP Coordination Mechanisms on EMS Coordination Point One 180 5.11 Average coordination cost of selected GPGP Coordination Mechanisms on EMS Coordination Point One 181 5.12 Average response time of selected GPGP Coordination Mechanisms on EMS Coordination Point Two 182 5.13 Average survival rate of selected GPGP Coordination. .. dispatch for an incident until reply messages arrive Furthermore, the task of a response agent can not start until the notification of the the selection of certain agents Clearly the dependency is between the message-sending task of the dispatcher and the task of a response agent 16 Figure 1.3: EMS coordination point two: between a police agent and an ambulance The second example is named EMS coordination. .. professor informs the assistant of an EST/deadline and gives the assistant a certain freedom to schedule his own activities as long as the assistant respects the EST/deadline Let us briefly discuss the naming of successor and predecessor here In simple words, a successor is an enablee agent that needs to request assistance for a coordination purpose from another agent; a predecessor is an enabler agent that