Online heuristic planning for highly uncertain domains

摘要

Heuristic search algorithms for online POMDP planning have shown great promise in creating successful policies for maximizing agent rewards using heuristics typically focused on reducing the error bound in the agent's cumulative future reward estimations. However, error bound-based heuristics are less informative in highly uncertain domains requiring long sequences of information gathering, such as robotics. In these domains, all possible plan improvements look similar under error bound-based heuristics until the agent's belief uncertainty has been resolved, leaving the agent initially confused on how best to improve its plan under the real-time constraints of online planning. We propose (1) a novel heuristic guiding the agent towards policies that first reduce the agent's belief uncertainty, after which error bound-based heuristics are more effective, and (2) a novel selection mechanism for choosing which type of heuristic (error bound or uncertainty-based) to use during the current stage of planning to most quickly form a good plan. We evaluate our solution in several benchmark POMDP problems, demonstrating that our solution yields successful policies with less planning time in highly uncertain domains and comparable performance in simpler problems.