Planning under uncertainty through goal-driven action selection

Abstract

Online planning in domains with uncertainty and partial observability conveys a series of performance challenges: agents must obtain information about the environment, quickly select actions with high reward prospects and avoid very expensive mistakes, while interleaving planning and execution in highly variable and uncertain domains. In order to reduce the amount of mistakes and help an agent focus on directly relevant actions, we propose a goal-driven, action selection method for planning in (PO)MDP’s. This method introduces a reward bonus and a rollout policy for MCTS planners, both of which depend almost exclusively on a clear specification of the goal and produced promising results when planning in large domains of interest to cognitive and mobile robotics.