Adapting Environment-Mediated Self-Organizing Emergent Systems by Exception Rules Holger Kasinger, Bernhard Bauer, Jörg Denzinger and Tom Holvoet
Programming Distributed Systems Lab Introduction Environment-mediated self-organizing emergent systems › Many, simple elements (mostly realized by agents) › Decision making solely based on locally available information › Local actions and interactions achieve global system goals › Usage of decentralized coordination mechanism »Pheromone-based coordination »Infochemical-based coordination »Field-based coordination Potential risk › Efficiency (performance) during operation cannot be guaranteed due to several runtime insufficiencies (system characteristics) June 7, Holger Kasinger
Programming Distributed Systems Lab Agenda Introduction Runtime insufficiencies Efficiency Improvement Advisor Exception rules Conclusions June 7, Holger Kasinger
Programming Distributed Systems Lab Runtime insufficiencies Case study: Dynamic Pickup and Delivery Problem (PDP) June 7, Holger Kasinger Task:
Programming Distributed Systems Lab Runtime insufficiencies Pollination-inspired coordination (PIC) for solving PDPs June 7, Holger Kasinger Task: Allomone Synomone Pheromone
Programming Distributed Systems Lab Runtime insufficiencies Field-based task assignment (FiTA) for solving PDPs June 7, Holger Kasinger
Programming Distributed Systems Lab Runtime insufficiencies Insufficiency 1: Reactiveness of agents June 7, Holger Kasinger
Programming Distributed Systems Lab Runtime insufficiencies Insufficiency 2: Greediness of agents June 7, Holger Kasinger
Programming Distributed Systems Lab Runtime insufficiencies Insufficiency 3: Absence of global knowledge June 7, Holger Kasinger
Programming Distributed Systems Lab Runtime insufficiencies Insufficiency 4: Inability to ‘look into future’ June 7, Holger Kasinger
Programming Distributed Systems Lab Agenda Introduction Runtime insufficiencies Efficiency Improvement Advisor Exception rules Conclusions June 7, Holger Kasinger
Programming Distributed Systems Lab Efficiency Improvement Advisor Specific constraints for feedback control loops › Low observability and poor controllability › Capability for self-organization and emergence › Openness and autonomy Assumptions and premises › Each agent is able to collect data about its local behavior › Each agent can be extended to a rule-applying agent › A sequence of runs (days) must have a (sub)set of similar tasks in (nearly) each run (day) June 7, Holger Kasinger
Programming Distributed Systems Lab Efficiency Improvement Advisor Functional architecture June 7, Holger Kasinger Receive histories Transform histories Extract recurring tasks 3 Optimize solution 4 Derive advice 5 (A,1) (B,2) (C,2) (D,3) (A,1) (B,2) (C,3) (D,2) Send advice 6 Centralized feedback control loop MAS
Programming Distributed Systems Lab Agenda Introduction Runtime insufficiencies Efficiency Improvement Advisor Exception rules Conclusions June 7, Holger Kasinger
Programming Distributed Systems Lab Exception rules Classification June 7, Holger Kasinger Exception rules Task-triggered rules Ignore rules Boost rules Wait rules Time-triggered rules Forecast rules Detection rules Neighborhood- triggered rules Idle rules Path rules Event-condition-action rules
Programming Distributed Systems Lab Exception rules Effect of ignore rules June 7, Holger Kasinger
Programming Distributed Systems Lab Exception rules Experimental evaluation June 7, Holger Kasinger Unoptimized solutionOptimized solution
Programming Distributed Systems Lab Exception rules Experimental results June 7, Holger Kasinger Random Scenarios Time Windows Changing Tasks 14% 17% Improvement
Programming Distributed Systems Lab Agenda Introduction Runtime insufficiencies Efficiency Improvement Advisor Exception rules Conclusions June 7, Holger Kasinger
Programming Distributed Systems Lab Conclusions Conclusions › EIA and exception rules »Adapt the local behavior of single agents in self-organizing systems »Improve the efficiency of the global solution »Take into account specific system constraints »More than just parameter adaptation › Current state »Done: ignore rules »In progress: boost rules, forecast rules »To be done: wait rules, detection rules, idle rules, path rules › Limitations »Not appropriate for problems without recurring tasks »Still limited in the size of problems June 7, Holger Kasinger
Programming Distributed Systems Lab Conclusions Open questions › How to guarantee that the adaptation of the local behavior is not counterproductive and possibly worsens the global solution in awkward situations? › Will scalability in terms of millions of agents be an issue for real-world application domains? › What is the trade-off for decentralizing the EIA approach in terms of additional communication and coordination efforts? › Assumed that an optimal solution to a problem can be calculated, how close can we get to this solution by an adapted self-organizing emergent system? June 7, Holger Kasinger
Programming Distributed Systems Lab Thank you for your attention! June 7, 2010Holger Kasinger22