Project: A Neuro-Symbolic Approach to Explainable Reinforcement Learning

Key information:

Abstract: 

Despite significant progress in deep reinforcement learning over the last 5 years across a range of environments, there are still limited tools for understanding how or why neural networks make certain decisions. In particular, how will an action impact future goals or aims of an agent? Answering questions such as this is important as reinforcement learning agents never take an action purely based on the current observation but in the context of a plan of how it will act in the future. A significant majority of previous research in explainable reinforcement learning has focused on the individual decision-making process an agent takes, often missing the temporal context that the agent acts in. In this research, we explore temporal explanation algorithms using a plan-based model of an agent to explain the future aims and goals of a sequence of actions. In this way, explanations may capture the intent behind specific action choices that observation-focused algorithms miss.