Evolutionary Learning of Interpretable Decision Trees

Reinforcement learning techniques achieved human-level performance in several tasks the last decade. However, recent years, need for interpretability emerged: we want to be able understand how a system works and reasons behind its decisions. Not only assess safety of produced systems, also it extract knowledge about unknown problems. While some that optimize decision trees reinforcement do exist, they usually employ greedy algorithms or not exploit rewards given by environment. This means these may easily get stuck local optima. In this work, propose novel approach interpretable uses trees. We present two-level optimization scheme combines advantages evolutionary with Q-learning. way decompose problem into two sub-problems: finding meaningful useful decomposition state space, associating an action each state. test proposed method on three well-known benchmarks, which results competitive respect state-of-the-art both interpretability. Finally, perform ablation study confirms using gives boost non-trivial environments one-layer technique.