Relevance-based Online Planning in Complex POMDPs

Planning under uncertainty is a central topic at the intersection of disciplines such as artificial intelligence, cognitive science and robotics, and its aim is to enable artificial agents to solve challenging problems through a systematic approach to decision-making. Some of these challenges include generating expectations about different outcomes governed by a probability distribution and estimating the utility of actions based only on partial information. In addition, an agent must incorporate observations or information from the environment into its deliberation process and produce the next best action to execute, based on an updated understanding of the world. This process is commonly modeled as a POMDP, a discrete stochastic system that becomes intractable very quickly. Many real-world problems, however, can be simplified following cues derived from contextual information about the relative expected value of actions. Based on an intuitive approach to problem solving, and relying on ideas related to attention and relevance estimation, we propose a new approach to planning supported by our two main contributions: PGS grants an agent the ability to generate internal preferences and biases to guide action selection, and IRE allows the agent to reduce the dimensionality of complex problems while planning online. Unlike existing work that improves the performance of planning on POMDPs, PGS and IRE do not rely on detailed heuristics or domain knowledge, explicit action hierarchies or manually designed dependencies for state factoring. Our results show that this level of autonomy is important to solve increasingly more challenging problems, where manually designed simplifications scale poorly