A co-evolutionary approach to interpretable reinforcement learning in environments with continuous action spaces

Machine learning (ML) has lately achieved impressive breakthroughs in several fields, enabling a plethora of exciting applications. However, mainstream ML techniques often have an undesirable property: they are not directly understandable by humans, thus humans cannot trust them in high-stakes or life-critical scenarios. A subfield of AI called interpretable AI (IAI) addresses this problem: generating models that are easy to understand for humans and, consequently, trustworthy. While several approaches apply IAI techniques to reinforcement learning problems, addressing the case in which an agent has to act in a continuous action space is still an open question. In this work, we propose a cooperative co-evolutionary approach based on grammatical evolution, Q-learning, and the univariate marginal distribution algorithm, specifically designed to train IAI agents (in the form of binary decision trees) capable of acting in environments with continuous action spaces. The experimental results show that our method is able to solve two well-known OpenAI Gym test cases reaching state-of-the-art performance. Moreover, a quantitative post hoc analysis reveals that the obtained solutions are more interpretable than those reported in the literature.