Pruning Decision Trees for Interpretable Reinforcement Learning

Researchers have introduced a pruning process to simplify decision-tree policies in reinforcement learning, aiming to enhance interpretability while preserving performance. The method, detailed in a recent arXiv preprint, defines structural and usage-aware operators to evaluate candidate edits by re-executing the policy, measuring return and interpretability proxies. This approach transforms complex policy structures into more compact ones, making edits auditable. The study applies the method to classic control and MuJoCo benchmarks, demonstrating consistent interpretability improvements with maintained high performance. The work addresses the challenge of inspecting reinforcement learning policies, which is crucial for trustworthiness. The preprint, titled "Interpretable reinforcement learning with decision-tree pruning," is available on arXiv under the Computer Science > Machine Learning category. The research contributes to the growing field of explainable AI, offering a practical tool for creating transparent AI systems.