Explaining the Influence of Prior Knowledge on POMCP Policies

Partially Observable Monte Carlo Planning is a recently proposed online planning algorithm which makes use of Monte Carlo Tree Search to solve Partially Observable Monte Carlo Decision Processes. This solver is very successful because of its capability to scale to large uncertain environments, a very important property for current real-world planning problems. In this work we propose three main contributions related to POMCP usage and interpretability. First, we introduce a new planning problem related to mobile robot collision avoidance in paths with uncertain segment difficulties, and we show how POMCP performance in this context can take advantage of prior knowledge about segment difficulty relationships. This problem has direct real-world applications, such as, safety management in industrial environments where human-robot interaction is a crucial issue. Then, we present an experimental analysis about the relationships between prior knowledge provided to the algorithm and performance improvement, showing that in our case study prior knowledge affects two main properties, namely, the distance between the belief and the real state, and the mutual information between segment difficulty and action taken in the segment. This analysis aims to improve POMCP explainability, following the line of recently proposed eXplainable AI and, in particular, eXplainable planning. Finally, we analyze results on a synthetic case study and show how the proposed measures can improve the understanding about internal planning mechanisms.