Pruning Decision Trees Improves RL Policy Interpretability

Reinforcement learning (RL) policies, despite their effectiveness, often lack transparency, making them difficult to trust and inspect. A common approach to enhance interpretability is to convert trained policies into explicit decision-tree rules. However, even these rule-based representations can remain overly complex for human understanding. This research proposes a novel pruning process specifically designed to simplify such decision-tree policies. The process employs a small set of structural and usage-aware operators to evaluate potential edits. Candidate edits are assessed by re-executing the policy to measure both task return and interpretability proxies, creating an auditable transformation path from intricate to compact policy structures. The effectiveness of this approach was demonstrated on classic control and MuJoCo benchmarks, where pruning consistently improved interpretability while preserving high performance.