Monte Carlo Tree Search for concurrent actions

Abstract

Many decision problems within robotics concern the control of potentially concurrent actions in continuous time, each with stochastic durations. There exist various formalizations of such decision processes. This thesis aims to investigate in a reduction that is suitable for Monte-Carlo Tree Search. In particular, the approach should be compared to existing reductions as Semi-Markov Decision Processes w.r.t. the generality of the formalisms as well as the notions of optimality guaranteed. Do the optimality proofs of Upper Confidence Bounds applied to Trees directly transfer to the concurrent action case? The handling of general stochasticity, also of the action durations, should be investigated in detail. Further, it is of interest to theoretically investigate and compare existing implementations w.r.t. the consistency and performance.