Presentation is loading. Please wait.

Presentation is loading. Please wait.

Adaptive Optimization of Solution Time In A Distributed Multi-agent System Amy Fedyk, Gary Kratkiewicz, Jeff Berliner, Mark Davis, Beth DePass, Rich Lazarus,

Similar presentations


Presentation on theme: "Adaptive Optimization of Solution Time In A Distributed Multi-agent System Amy Fedyk, Gary Kratkiewicz, Jeff Berliner, Mark Davis, Beth DePass, Rich Lazarus,"— Presentation transcript:

1 Adaptive Optimization of Solution Time In A Distributed Multi-agent System Amy Fedyk, Gary Kratkiewicz, Jeff Berliner, Mark Davis, Beth DePass, Rich Lazarus, Rusty Bobrow KIMAS, April 18, 2005 Afedyk@bbn.com

2 2Outline Optimization goalOptimization goal UltraLog OverviewUltraLog Overview Prior ArtPrior Art Solution Time Optimization ChallengesSolution Time Optimization Challenges Techniques for Optimizing Solution TimeTechniques for Optimizing Solution Time ConclusionConclusion

3 3 Optimization Goal Improve time to solution in a large-scale logistics planning applicationImprove time to solution in a large-scale logistics planning application –Have a solution available at all times –Eliminate unnecessary re-work –Minimize effects of perturbations within society –Continue to perform during system stresses and communications loss

4 4 UltraLog: A Large Agent Society UltraLog UltraLog – DARPA-funded effort to explore building logistics systems with a distributed multi-agent architecture – The test society models demand from military organizations supported by a logistics supply chain Each agent models a single military organization with its physical assets, business rules, and relationships to other organizations Each agent models a single military organization with its physical assets, business rules, and relationships to other organizations Contains over 1000 medium weight agents distributed across nearly 100 computers Contains over 1000 medium weight agents distributed across nearly 100 computers –Built with Cougaar Open source, distributed-agent architecture Open source, distributed-agent architecture

5 5 Solution Time Optimization Challenges Large-scale military logistics planning application Small changes can affect many agents within the society.Small changes can affect many agents within the society. Supporting agents do not know when all their requests have been received.Supporting agents do not know when all their requests have been received. 1-AD 16-CSG (1-AD) 123- MSB-POL 102-POL- SUPPLYCO OSD USAEUR USEUCOM 5-CORPS REAR 7-CSG (5-CORPS) 240- SSCO 3-SUPCOM -HQ DLAHQOSCTRANSCOM 21-TSC -HQ 110-POL- SUPPLYCO HNS 5-CORPS ARTY FORSCOM 1-AD Orgs 16-CSG Orgs 7-CSG Orgs 5-CORPS REAR Orgs 5-CORPS ARTY Orgs 21-TSC Orgs 26- SSCO 900-POL- SUPPLYCO 574- SSCO 3-SUPCOM Orgs Fuel Supply Requests Fuel Supply Chain

6 6 Prior Art Adaptive systems Adaptive systems – Gracefully degrading systems – Survivable systems – Self-healing systems – Speculative computation Effects of communication on performance Effects of communication on performance – Trade-off cost of communication and value of information Building on prior artBuilding on prior art – “Self-pacing” system – Graceful degradation via speculative computation – Improve performance by limiting information flow in a purposeful manner

7 7 Techniques For Optimizing Solution Time 1.Multi-Resolutional solutions –Continuous up-to-date plan –Adapts to system stresses 2.Control upward/downward information flow –Propagate change based on local consistency 3.Transmission of differences only –Each agent minimizes effects of changes by transmitting only the differences between the previously seen and new plan 4.Use predictors –Proxies for temporarily unavailable components

8 8 1. Multi-Resolutional Solutions Society generates two plans simultaneouslySociety generates two plans simultaneously Low-resolution solutionLow-resolution solution –Rough estimate plan –Produced quickly –Preferred over no solution High-resolution solutionHigh-resolution solution –Detailed high fidelity plan –Becomes available more slowly –Gradually replaces low-resolution solution –Allows the plan to evolve over time

9 9 Replace Low for High-Resolution The high-resolution solution gradually replaces the low-resolution solution Initial Solution Low Ultimate Solution High Near-Term TasksLong-Term Tasks Time elapsed while planning Still Better Solution Low High Low Better Solution HighLow HighLow

10 10 2. Controlling Upward/Downward Information Flow Information Flow in the Supply Chain 1.Local Agent receives incoming tasks from customers 2.Local Agent sends outgoing messages to providers. 3.Local Agent receives responses back from providers 4.Local Agent then sends responses back to its customers

11 11 2. Controlling Upward/Downward Information Flow Reduce solution time by managing re-workReduce solution time by managing re-work –Local agents refrain from sending messages if local re- work is likely Incoming tasks have changedIncoming tasks have changed Greatly improved stability and performance.Greatly improved stability and performance. –Test societies of 1092 agents show solution times which are always under 12 minutes on baseline runs.

12 12 3. Transmit Differences Only Minimize the affects of perturbations.Minimize the affects of perturbations. Each agent evaluates the messages it has previously sent to its providers before sending the re-computed plan.Each agent evaluates the messages it has previously sent to its providers before sending the re-computed plan. Transmission-of-differences technique reduced number of unnecessary perturbations in society by an average of 26.0%.Transmission-of-differences technique reduced number of unnecessary perturbations in society by an average of 26.0%. Transmit only the one changed task Transmit only two changed tasks and responses

13 13 4. Predictors Predictors are agent proxies which provide approximations based on the best available data.Predictors are agent proxies which provide approximations based on the best available data. The predictors allow agents to continue planning during comms lossThe predictors allow agents to continue planning during comms loss –Customer Predictors (CP) estimate incoming customer requests. –Supplier Predictor (SP) estimates answers a supplier would give in response to customer requests. Under loss of comms, agents with predictors were about 3x faster than agents without predictors.Under loss of comms, agents with predictors were about 3x faster than agents without predictors. Customer Agent Supplier Agent Customer Agent Customer Agent SP CP Customers estimate a supplier’s response Supplier estimates a customer’s requests

14 14Conclusion Multi-Resolutional solutions provide a continuously available and continuously improving planMulti-Resolutional solutions provide a continuously available and continuously improving plan Controlling Upward/Downward information flow prevents unnecessary re-work.Controlling Upward/Downward information flow prevents unnecessary re-work. Exclusive transmission of differences minimizes effects of perturbations.Exclusive transmission of differences minimizes effects of perturbations. Predictors allow computation to proceed during comms loss.Predictors allow computation to proceed during comms loss.

15 15 For more information … BBN Technologies:BBN Technologies: –http://www.bbn.com http://www.bbn.com Cougaar Agent Architecture:Cougaar Agent Architecture: –http://www.cougaar.org http://www.cougaar.org Other Cougaar-related KIMAS’05 papers:Other Cougaar-related KIMAS’05 papers: – “ Watching Your Own Back: Self Managing Multi-Agent Systems ”, M. Thome, T. Wright, et al – “, J. Zinky, S. Siracuse, et al – “ Using QoS-Adaptive Coordination Artifacts to Increase Scalability of Communication in Distributed Multi-Agent Systems”, J. Zinky, S. Siracuse, et al – –“A Reconfigurable Multiagent Society for Transportation Scheduling and Dynamic Rescheduling”, D. Montana, G.Vidaver, et al – –“Scalability Aspects of Agent-based Naming Services”, T. Wright and K. Kleinmann


Download ppt "Adaptive Optimization of Solution Time In A Distributed Multi-agent System Amy Fedyk, Gary Kratkiewicz, Jeff Berliner, Mark Davis, Beth DePass, Rich Lazarus,"

Similar presentations


Ads by Google