Multi-objective Genetic Programming for Explainable Reinforcement Learning

Deep reinforcement learning has met noticeable successes recently for a wide range of control problems. However, this is typically based on thousands of weights and non-linearities, making solutions complex, not easily reproducible, uninterpretable and heavy. The present paper presents genetic programming approaches for building symbolic controllers. Results are competitive, in particular in the case of delayed rewards, and the solutions are lighter by orders of magnitude and much more understandable.