Bounty Hunting and Multi- Robot Task Allocation. Drew Wicke

Transcription

1 Bounty Hunting and Multi- Robot Task Allocation Drew Wicke

2 Outline Introduction Auction Based Task Allocation Bounty Hunting based Task Allocation Dynamic Tasks Results Task Abandonment Results Future Work Multi-robot task allocation Cooperative tasks Large scale 2

3 Decentralized Dynamic Task Allocation Agents are mapped to tasks which they must complete Agents decide collectively how to map the tasks New tasks appear dynamically while agents are working on other tasks Real Life Applications Disaster and Emergency Tasks Robot Soccer Pickup and Delivery problems Common Methods Auctions Markets Contract Networks Dias, M. B., Zlot, R., Kalra, N., & Stentz, A. (2006). Market-based multirobot coordination: A survey and analysis. Proceedings of the IEEE, 94(7),

4 Taxonomy of Task Allocation Single-Task Robots vs Multi-Task Robots Single Robot Tasks vs Multi-Robot Tasks Instantaneous Assignment vs Time Extended Gerkey, B. P., & Matarić, M. J. (2004). A formal analysis and taxonomy of task allocation in multi-robot systems. The International Journal of Robotics Research, 23(9),

5 Auctions (MURDOCH) Bid valuation (or fitness score) for each task Auctioneer determines winner after some clearing time Auctioneer monitors winners and reassigns if insufficient progress 5

6 Auctions are a strange fit Auctions are generally exclusive, assume infinite money supply and altruistic, nonstrategic agents, require accurate valuation, and commonly assume sequential task arrival or some kind of task queue. This is an unwieldy and unreasonable set of constraints for many dynamic task allocation problems. 6

7 An alternative: Bounty Hunting! Bounty hunter and Bail bondsman Task Classes, and task commitment Advantage: no bidding! Disadvantage: Not exclusive Drew Wicke and David Freelan and Sean Luke (2015, May). Bounty Hunters and Multiagent Task Allocation. In Proceedings of the 2015 AAMAS (pp ). 7

8 Problem Description 8

9 Experiments Static Environment Dynamic Agents - agents are removed Dynamic Tasks - task distribution changes Unreliable Collaborators - two slow agents are added, pick tasks randomly. Unexpectedly Bad Tasks - randomly selected tasks become difficult for some agents Evaluated by measuring the total bounty available. 9

10 Bounty Hunters Bounty hunters learn which task classes to go after Simple - learns time to task and probability of success Complex - learns time to task and probability of success given that certain other agents have committed to the task. Bounty Auction Similar to an Auction, but incorporates the Bounty 10

11 Results Adaptive bounties converge to divvying tasks up despite not being exclusive Adaptive Bounties can successfully adapt to dynamically changing environments Adaptive Bounties can perform comparably to, and in some cases much better than, exclusive methods Complex is generally more effective 11

12 Task Abandonment Bounty hunters may abandon tasks, or Jumpship. Improves performance significantly in all but the Dynamic Agents and Dynamic Tasks experiments. Does statistically significantly the best in two new experiments. Wicke, D., Wei, E., & Luke, S. Throwing in the Towel:Faithless Bounty Hunters as a Task Allocation Mechanism. Submitted AAMAS

13 Multi-robot Bounty Hunting Current assumptions Teleportation to home base after task completion Home Base Simultaneous occupation of space Grid movement (manhattan distance) Instead Full pickup and delivery Obstacle avoidance Continuous space (cartesian distance) Incorporate cost and battery life 13

14 Future Work Implement Bounty Hunting system in a multirobot system using GMU s FlockBots. Important to show that the system works in real life, not just in simulation. 14

15 Bounty Hunting and Multi-Robot Task Allocation Dias, M. B., Zlot, R., Kalra, N., & Stentz, A. (2006). Market-based multirobot coordination: A survey and analysis. Proceedings of the IEEE, 94(7), Gerkey, B. P., & Matarić, M. J. (2004). A formal analysis and taxonomy of task allocation in multi-robot systems. The International Journal of Robotics Research, 23(9), Gerkey, B. P., & Matari, M. J. (2002). Sold!: Auction methods for multirobot coordination. Robotics and Automation, IEEE Transactions on, 18(5), Wicke, D., Freelan, D., & Luke, S. (2015, May). Bounty Hunters and Multiagent Task Allocation. In Proceedings of the 2015 International Conference on Autonomous Agents and Multiagent Systems (pp ). International Foundation for Autonomous Agents and Multiagent Systems. Wicke, D., Wei, E., & Luke, S. Throwing in the Towel:Faithless Bounty Hunters as a Task Allocation Mechanism. Submitted AAMAS

16 Bounty Hunting and Multi- Robot Task Allocation Questions 16