Time-adaptive cross entropy planning

Abstract

We propose Time-Adaptive Cross Entropy Planning (Tace) to increase flexibility of online planning agents in continuous state, action and time domains with infinite state-action spaces and branching factors. Tace reduces simulation effort of planning with cross entropy optimization by maintaining and adapting a probability distribution over the optimal planning horizon. This allows to identify temporally local problems in a global context, and to subsequently concentrate on the solution of the local problem. We show the effectiveness of Tace by comparing it empirically to a state-of-the-art online planner for continuous domains.