Output details
11 - Computer Science and Informatics
University College London
Monte-Carlo Planning in Large POMDPs
<12>Stochastic, partially observable decision-making problems (POMDPs) are considered the hardest class of planning problems. The POMCP algorithm, developed in this paper, was the first general, high-performance planning algorithm for this class of problem. It equalled the state-of-the-art on existing benchmark problems, but was also able to achieve good performance on problems that were orders of magnitude larger (e.g. one problem had 10^60 states). POMCP was proven to converge to the optimal planning solution, with asymptotic bounds on the rate of convergence. Wu, Lee and Hsu used POMCP to win the partially observable track of the 2011 International Probabilistic Planning Competition.