Multiagent Planning with Bayesian Nonparametric Asymptotics

Autonomous multiagent systems are beginning to see use in complex, changing environments that cannot be completely specified a priori. In order to be adaptive to these environments and avoid the fragility associated with making too many a priori assumptions, autonomous systems must incorporate some form of learning. However, learning techniques themselves often require structural assumptions to be made about the environment in which a system acts. Bayesian nonparametrics, on the other hand, possess structural flexibility beyond the capabilities of past parametric techniques commonly used in planning systems. This extra flexibility comes at the cost of increased computational cost, which has prevented the widespread use of Bayesian nonparametrics in realtime autonomous planning systems. This thesis provides a suite of algorithms for tractable, realtime, multiagent planning under uncertainty using Bayesian nonparametrics. The first contribution is a multiagent task allocation framework for tasks specified as Markov decision processes. This framework extends past work in multiagent allocation under uncertainty by allowing exact distribution propagation instead of sampling, and provides an analytic solution time/quality tradeoff for system designers. The second contribution is the Dynamic Means algorithm, a novel clustering method based upon Bayesian nonparametrics for realtime, lifelong learning on batch-sequential data containing temporally evolving clusters. The relationship with previous clustering models yields a modelling scheme that is as fast as typical classical clustering approaches while possessing the flexibility and representational power of Bayesian nonparametrics. The final contribution is Simultaneous Clustering on Representation Expansion (SCORE), which is a tractable model-based reinforcement learning algorithm for multimodel planning problems, and serves as a link between the aforementioned task allocation framework and the Dynamic Means algorithm. Thesis Supervisor: Jonathan P. How Title: Richard C. Maclaurin Professor of Aeronautics and Astronautics