DECENTRALIZED MULTIAGENT METAREASONING APPLICATIONS IN TASK ALLOCATION AND PATH FINDING

Abstract

Decentralized task allocation and path finding are two problems for multiagent systems where no single fixed algorithm provides the best solution in all environments. Past research has considered metareasoning approaches to these problems that take in map, multiagent system, or communication information. None of these papers address the application of metareasoning about individual agent state features which could decrease communication and increase performance for decentralized systems. This thesis presents the application of a meta-level policy that is conducted offline using supervised learning through extreme gradient boosting. The multiagent system used here operates under full communication, and the system uses an independent multiagent metareasoning structure. This thesis describes research that developed and evaluated metareasoning approaches for the multiagent task allocation problem and the multiagent path finding problem. For task allocation, the metareasoning policy determines when to run a task allocation algorithm. For multiagent path finding, the metareasoning policy determines which algorithm an agent should use. The results of this comparative research suggest that this metareasoning approach can reduce communication and computational overhead without sacrificing performance.