Transcript Kapitel 13
Universität Karlsruhe (TH) Flexibility Through Multiagent Systems: Solution or Illusion? Peter C. Lockemann © 2004 IPD, Prof. Lockemann SOFSEM04 Scenario: Production and supply chain Monitoring Execution Planning ‹#› Assembly planning Dispoweb Global planning and negotiations Production KRASHplanning for MAS mechanics IntaPSFABMASMAS MAS Production planning for electronics intraplant © 2004 IPD, Prof. Lockemann Customer OEM dispoweb dispoweb ATT/SCC interplant Tracking external SOFSEM04 Scenario: Production and supply chain ‹#› DISPOWEBPlaning MAS IntaPSExe- MAS cution Tracing DISPOWEB- Tracing MAS Planing ATTMAS SCCMAS Execution KRASH-MAS ATT-MAS ATTMAS Tracing Execution FABMAS DISPOWEBMAS Planing Negotiation of the global production schedule Tracing of processes and propagation of delays © 2004 IPD, Prof. Lockemann SOFSEM04 Scenario: Production and supply chain ‹#› Nr. Activity (Actor) Negotiate initial plan of production between supply chain partners (DISPOWEB). Operational assembly planning (KRASH). Production planning for e.g. mechanical parts (IntaPS). Production planning for e.g. electronic parts (FABMAS). Monitoring of orders and related suborders (ATT). Trigger internal planning systems in case of minor critical events (ATT). Next Trigger re-planning by DISPOWEB agents in case of a severe critical event (ATT). Next Controlling information is forwarded to trusted third party SCC-system (ATT/SCC). Internal rescheduling in reaction to a critical event (KRASH, IntaPS, FABMAS). Next Renegotiate plan of production between supply chain partners due to severe critical event (DISPOWEB). Next © 2004 IPD, Prof. Lockemann SOFSEM04 Scenario: Health care ‹#› Kliniken & Stationen Internal Medicine Surgery Ward Uni Hohenheim/ Hosp. Sindelf inden (Referring) Physicians TU Ilmenau/ Phy sicians Thuringia, OnkoNet N.N. Uni Mannheim/ Hosp. Pegnitz, Hosp. Kulmbach (…/…) Reception Ward Uni Hohenheim/ Hosp. Sindelf ingen Gastro-Enterology Cancerology Uni Würzburg/ Univ .-Hosp. Würzburg TU Ilmenau/ Univ .-Hosp. KIM II Jena Radiology & Radiotherapy Conv., CT, MRT Cardiological Laboratory Uni Mannheim/ Hosp. Pegnitz & Hosp. Kulmbach Uni Mannheim/ Hosp. Pegnitz, Hosp. Kulmbach Transport Serv. Uni Freiburg/ Univ .-Hosp. Freiburg Radiotherapy Uni Würzburg/ Univ .-Hosp. Würzburg Radiodiagnosis Uni Freiburg/ Univ .-Hosp. Freiburg N.N. (…/…) Operating Theatre Surgery Block Ambulance TU Ilmenau/ Ambulances Thuringia Emergency Dept. TU Berlin/ Hosp. Charité Berlin Univ Trier/Hosp. Barmh. Brüder Trier Dept. of Surgery Uni Mannheim/ Hosp. Pegnitz & Hosp. Kulmbach © 2004 IPD, Prof. Lockemann Op.-Planning TU Berlin/ Charité Berlin N.N. Intensive Care Unit TU Berlin/ Hosp. Charité Berlin (…/…) SOFSEM04 Scenario: Digital libraries ‹#› customer user agent trader wrapper order and delivery provider HTTP queries to wrappers and query results HTTP HTTP query registration and result presentation queries to traders and trader recommendations provider registration HTTP The UniCats environment © 2004 IPD, Prof. Lockemann SOFSEM04 Scenario: Traffic prognosis ‹#› © 2004 IPD, Prof. Lockemann SOFSEM04 Scenario: Traffic prognosis ‹#› © 2004 IPD, Prof. Lockemann SOFSEM04 The optimist ‹#› Social organizations are resilient and responsive to ever changing situations because they include enough intelligent decision makers sufficient slack to take decisions. Then why not mirror these properties to build software systems that are flexible (resilient and responsive)? Agents are a most natural metaphor to model complex social organizations that must be highly adaptive and rely on flexible and adaptive members. Such systems are known from AI and referred to as multiagent systems (MAS). Hypothesis: If well done, MAS offer the adaptivity and flexibility needed for the modeled world. © 2004 IPD, Prof. Lockemann SOFSEM04 The pessimist ‹#› Multiagent systems are distributed include asynchronous communication have components with indeterministic behavior. Therefore, they are inherently difficult to engineer: It is hard to give a sufficiently precise system specification and even if there is one, what are the chances to verify or validate that the system implementation satisfies the specification? © 2004 IPD, Prof. Lockemann SOFSEM04 Where the two may meet ‹#› Are there situations with compelling reasons for a multiagent approach? Can one yet impose constraints to be able to engineer multiagent systems? © 2004 IPD, Prof. Lockemann SOFSEM04 Agenda ‹#› Are there situations with compelling reasons for a multiagent approach? What is an agent after all? ... and a multiagent system? Useful multiagent software systems: A hypothesis Testing the hypothesis Mixed results and a refined hypothesis Can one yet impose constraints to engineer MAS? A database person’s view At least it should be reliable! Transactional agents Reliable agent cooperation Conclusions and challenges © 2004 IPD, Prof. Lockemann SOFSEM04 What is an agent? ‹#› Wooldrige and Jennings on agents: “An agent is a computer system that is situated in some environment, and that is capable of autonomous action in its environment in order to meet its design objectives.“ Wooldrige on intelligent agents: “An intelligent agent is a reactive, proactive, interacting agent.“ Flexibility © 2004 IPD, Prof. Lockemann SOFSEM04 What’s behind the definition ‹#› A piece of code: software agent An open system: must have an observable effect Effective indeterminism in time and kind “An agent is a computer system that is situated in some environment, and that is capable of autonomous action in its environment in order to meet its design objectives.“ Agents can be many things to many people “An intelligent agent is a reactive, proactive, interacting agent.“ Stimulators and responders © 2004 IPD, Prof. Lockemann SOFSEM04 Multiagent systems ‹#› autonomous situated reactive proactive Individual objectives interacting autonomous situated reactive proactive interacting Commo n objectiv e autonomous situated reactive proactive interacting © 2004 IPD, Prof. Lockemann SOFSEM04 Multiagent systems ‹#› autonomous situated reactive proactive Individual objectives interacting autonomous situated reactive proactive Can the systems be engineered? © 2004 IPD, Prof. Lockemann interacting Commo n objectiv e interacting autonomous situated reactive proactive How much flexibility can we manage? SOFSEM04 Who needs all that flexibility? ‹#› Hypothesis: the database person’s view the software architect’s view Multiagent systems are useful for an environment where the range of situations (the problem space) is too large for enumeration and instead needs a qualitative description, the problem space can be divided into sets of simpler tasks, each requiring specialized competence, the simpler tasks can be dealt with autonomously by individual agents, solving the overall situation requires cooperation among the agents. the AI person’s view © 2004 IPD, Prof. Lockemann the programmer’s view SOFSEM04 Flexibility through multiagent systems ‹#› autonomous situated reactive proactive interacting autonomous situated reactive proactive interacting Assignment of individual responsibilities & cooperative problem solving autonomous situated reactive proactive interacting © 2004 IPD, Prof. Lockemann SOFSEM04 Testing the hypothesis: Scenario ‹#› Product Assembly1 UL01 UF01 Supplier Buffer1 Product Assembly2 Product Assembly3 UL02 Immidiate Buffer2 UF02 UG01 Immidiate Buffer UA01 UG02 UA02 B1 A1 C2 B2 A2 C3 B3 A3 C4 B4 A4 C5 B5 A5 UB01 Mold Buffer Option C1 UB02 Buffer Shipping UC01 UH01 Product Assembly Area © 2004 IPD, Prof. Lockemann Supplier Buffer2 Unit Assembly Area Mold Maintenance UC02 Buffer Molding Area SOFSEM04 Benchmark ‹#› Kanban A C Product assembly Molding area B Buffer Manual assembly High variability within External homogeneous product spectrum supplier Unit assembly Flexibility via Kanban production organization: Order Supply buffer minimization via pull principle Production control: © 2004 IPD, Prof. Lockemann SOFSEM04 Benchmark ‹#› External supplier Kanban A C Product assembly Molding area B Buffer Manual assembly Externally determined production schedule Unit assembly Analytical and simulation-based layout Order planning for a givenSupply production program © 2004 IPD, Prof. Lockemann SOFSEM04 Test and Benchmark ‹#› Centrally planned Kanban is not flexible enough to deal with machine failures short-term deviations from the delivery schedule Solution options Passive: Simulate disturbances and adjust layout planning Reactive: Rescheduling algorithms (known to be suboptimal) Reactive: Agents Comparison by simulation © 2004 IPD, Prof. Lockemann SOFSEM04 The best performing agent model ‹#› Order Data Order Agent Machine Agent 1: New Order 2: Request Mixed-model Assembly Line Balancing Problem (MALBP): Reactive MAS Approach 3: Task Announcement 4: Task Announcement Bidding 5: Bid 6: Bid Processing 7: Reporting Results © 2004 IPD, Prof. Lockemann SOFSEM04 Experiments by simulation ‹#› Parameter variations: Master data: Bill of materials Operation list fore each product Transport lot size Affects the number of suborders for a single order Production order generated by production planning and control Disturbance profile per machine (statistical) Interval between disruptions Disruption duration Output: Throughput: Average Standard deviation © 2004 IPD, Prof. Lockemann SOFSEM04 Experiments by simulation ‹#› Corresponding Standard Deviations Throughput and Processing Time Parameterisation © 2004 IPD, Prof. Lockemann SOFSEM04 Experiments: Sample Analysis ‹#› Throughput Time 5 3 Lot Size 1-1,5 0,5-1 1 5 min 30 min 60 min Suitability of MAS Compared to Centralized OR Approaches Depending on Decision Variables 0-0,5 Disruption Duration Throughput Time Improvement by MAS Standard Deviation 5 Reliability and Predictability of the Results 3 Lot Size 1-1,5 0,5-1 1 5 min 30 min 60 min 0-0,5 -1-0 Disruption Duration Standard Deviation © 2004 IPD, Prof. Lockemann SOFSEM04 Empirical This results cannot solely be explained by the original hypothesis ‹#› Multiagent systems need slack: If assembly lines run close to capacity, MAS are even inferior. They performed better if slack was provided by additional machines or by delaying assembly orders. Multiagent systems need inhomogeneity: More assembly lines by themselves have no effect unless the machines follow very different disturbance profiles. Different disturbance profiles have even a positive effect if no machines are added. Multiagent systems operate best in a dynamic, nondeterministic environment: MAS are superior when it comes to evening out fluctuations due to disruptions. © 2004 IPD, Prof. Lockemann SOFSEM04 Revisiting the hypothesis ‹#› Hypothesis: Multiagent systems are useful for an environment where the range of situations (the problem space) is too large for enumeration and instead needs a qualitative description, the range of decisions (the solution space) is commensurate in size with the problem space, the problem space can be divided into sets of simpler tasks, each requiring specialized competence, the simpler tasks can be dealt with autonomously by individual agents, solving the overall situation requires cooperation among the agents. © 2004 IPD, Prof. Lockemann SOFSEM04 Practical implementation ‹#› FIPA-OS Implementation eMPlant Simulation Model DB Communication Platform using Event Mechanisms © 2004 IPD, Prof. Lockemann SOFSEM04 A database persons’ view ‹#› Hypothesis: Take a descriptive approach! Multiagent systems After all, SQL does are it! useful for an environment where your objectives, let the determine way State the range of situations (thesystem problem space) isthe toobest large for how to achieve and them! enumeration instead needs a qualitative description, need software engineering for MAS. We the range of decisions (the solution space) is Disciplines that could help: commensurate in size with the problem space, Model-driven programming the problem space can be divided into sets of simpler Service-oriented tasks, each requiringprogramming specialized competence, Component-orientation the simpler tasks can be dealt with autonomously by Agent platforms are available. individual agents, But then, agents and MAS should not be indeterministic solving the overall situation requires cooperation among beyond their own control, such as own disturbances! the agents. © 2004 IPD, Prof. Lockemann SOFSEM04 Mastering the complexity ‹#› autonomous situated reactive proactive interacting autonomous situated reactive proactive interacting No uncontrolled disturbances! autonomous situated reactive proactive interacting © 2004 IPD, Prof. Lockemann SOFSEM04 Transactions ‹#› autonomous situated reactive proactive Transactional agents interacting autonomous situated reactive proactive Agent may be involved ininteracting interacting autonomous situated reactive proactive Transactional conversations concurrent conversations! © 2004 IPD, Prof. Lockemann SOFSEM04 INTERRAP BDI agent architecture ‹#› Agent knowledge base Cooperative planning layer Cooperation Model SG PS Mental Model SG PS Local planning layer World Model SG PS Behavior based layer World Interface SG: Situation recognition and Goal activation PS: Planning, Scheduling and execution information access control flow © 2004 IPD, Prof. Lockemann SOFSEM04 Corresponding transaction model ‹#› Recovery Open nested transactions action node control node synchronization node parallel execution sequential execution alternate execution Conversation © 2004 IPD, Prof. Lockemann SOFSEM04 Evolutionary transactions ‹#› Agent knowledge base Cooperative planning layer Cooperation Model SG PS Mental Model SG PS Local planning layer World Model SG PS Behavior based layer World Interface ... © 2004 IPD, Prof. Lockemann SOFSEM04 Agent cooperation by synchronization ‹#› agent 1 transaction © 2004 IPD, Prof. Lockemann agent 2 transaction SOFSEM04 Agent cooperation by synchronization ‹#› Independent start prevents termination of slave before termination of master Task Delegation M1 Master M11 M111 M12 M112 M121 M13 ... M131 ... Slave S0 M1111 ... S2 S1 M112 waits for S1 to regain control S11 M11 wakes up S11 © 2004 IPD, Prof. Lockemann S121 S12 S122 compensate in slave if M112 fails S123 SOFSEM04 Experiments with an orthogonal architecture ‹#› common goal domain agent domain agent cooperation local actions local actions distributed plan transactional agent transactional agent local TA tree coordination local TA tree action execution action execution Isolation robustness service action (DB transaction) database database © 2004 IPD, Prof. Lockemann SOFSEM04 Not a nice solution because … ‹#› Cooperation protocol directly built into individual agent transaction agent 1 transaction Need for a shared database for conflict resolution © 2004 IPD, Prof. Lockemann agent 2 transaction Asynchronous communication via ECA rules that must be coordinated with database, e.g. via active database SOFSEM04 Transactional conversations ‹#› agent 1 transaction agent 2 transaction conversation distributed transaction © 2004 IPD, Prof. Lockemann SOFSEM04 Modeling of conversations ‹#› initiator participant initiator ready cfp_ sent cfp refuse not-understood propose reject-proposal prop_ refused not_ under stood prop_ sent prop_ refused abort prop_ acceptd accept-proposal failure participant inform-done inform-ref © 2004 IPD, Prof. Lockemann deal_ reached SOFSEM04 Transactional conversations ‹#› Transaction manager (CORBA OTS) DB I TA Task User Task Agent A MAS Platform I © 2004 IPD, Prof. Lockemann DB II FIPA conformant conversation TA Task User Task Drawback: Conversations must be attached to the leaf nodes Agent B MAS Platform II SOFSEM04 The next idea ‹#› agent 1 transaction agent 2 transaction conversation Message-oriented middleware (MOM) with messages based on speech acts © 2004 IPD, Prof. Lockemann SOFSEM04 Conclusions: Flexibility ‹#› MAS are very costly to engineer: System-level behavior is close to impossible to predict analytically, requires large simulative or experimental effort. The added cost does not even pay off in the majority of cases! The expenses seem only justified for applications with large problem and solution spaces, where the problem space is or appears non-deterministic. A natural application seems health care. Candidate technical applications are those with a high level of disruptions. © 2004 IPD, Prof. Lockemann SOFSEM04 Conclusions: Flexibility ‹#› Approach: Transactional properties Requires an expensive and extensive infrastructure of database manager and transaction manager. Heavy-weight nodes needed for agents. Distributed transactions (a foe of autonomy!) and active database mechanisms (e.g., trigggers) needed. © 2004 IPD, Prof. Lockemann SOFSEM04 Challenges: Industrial strength ‹#› Transactional synchronization: Evolutionary transactions to deal with reaction. Transactions with built-in cooperation. Non-orthogonal individual and cooperative behavior. Transactional conversations: Difficult to attach to agent transactions. Rigid, must have ACID properties. Message-oriented middleware: Still an unknown area for agents. Long term: A combination of transactional agents and message-oriented middleware with higher-level guarantees. Contribute to standards! © 2004 IPD, Prof. Lockemann SOFSEM04 Conversation Policy XML (cpXML) ‹#› • Origin: • • • • V 1.0 (August 2002) • Specification: XML Schema, Standards document, papers, tutorial Roles State InitialState Coverage: Sequencing and timing of messages Tools • • • ConversationPolicy Status: • • IBM (Conversation Support for Webservices) Reference implementation for WebSphere JCA extension by Conversation Manager and Conversation Adapter Role SendMessageTransition TimeoutTransition LoadChild ChildReturnTransition Remarks • Compatible to BPEL4WS • CPs exchangeable © 2004 IPD, Prof. Lockemann SOFSEM04 FIPA-Standard ‹#› Applications •Nomadic Application Support •Agent Software Integration Agent Message Transport •Msg Transport Service •Messaging Interoperability ACL Representations •ACL Msg Representation in •Bit Efficient •String •XML Envelope Representations •Agent Msg Transport Envelope Representation in •XML •Bit Efficient Transport Protocols •Agent Msg Transport Protocol for •IIOP •WAP •HTTP © 2004 IPD, Prof. Lockemann •Message Buffering •QoS •Network Mgmt •Personal Assistant, Personal Travel Assistant •Audio-visual Entertainment and Broadcasting Abstract Architecture •Abstract Architecture •Domains and Policies Foundation of Interoperable Agents Agent Communication •Ontology Service •ACL Message Structure Interaction Protocols •IP Library Spec •Request, Request-When •Query •Subscribe •Propose •Contract Net, Iterated ~ •English, Dutch Auction •Brokering, Recruiting Agent Management Communicative Acts •Agent Mgmt Spec •CA Library Spec Content Languages •Content Lang. Spec •SL, CCL, KIF, RDF SOFSEM04 Augmenting FIPA ‹#› Order agent Machine agent Data Warehouse „Agent“ Standard system DW Agent Directory Management Facilitator System „Yellow Pages“ „White Pages“ Standard system ... ... Task Standard system Scheduling Task Task Task Data Warehouse Software Robustness service FIPA-OS © 2004 IPD, Prof. Lockemann SOFSEM04