Decision-Theoretic Foundations for Multi-Agent Systems.

摘要

This project produced efficient algorithms for planning and coordination of multi-agent systems that can cope with uncertainty and missing information. These algorithms employ new plan representations and dynamic programming techniques that can exploit heuristic knowledge and the structure of the problem to improve scalability. The project produced mechanisms that exploit randomization to improve coordination and minimize communication, and has shown how to use agent goals to develop bounded-optimal algorithms that are based on sampling. Additionally, the project produced CBDP, an efficient and scalable point-based dynamic programming algorithm for Network Distributed POMDPs, particularly suited for managing sensor network tracking tasks. A formal framework for decentralized monitoring has been developed for coordination of agents that solve components of a larger problem in a decentralized manner. These new coordination algorithms have been rigorously evaluated and shown to produce magnitudes of speedup in policy computation and better quality solutions than state-of-the-art methods.