Toward Interpretable-AI Policies Using Evolutionary Nonlinear Decision Trees for Discrete-Action Systems

Area of Science:

• Artificial Intelligence
• Control Theory
• Machine Learning

Background:

• Black-box AI, particularly deep reinforcement learning (DRL), is widely used for control tasks but lacks interpretability.
• DRL policies, while efficient, are often complex and difficult to understand or explain.

Purpose of the Study:

• To develop an interpretable Nonlinear Decision Tree (NLDT) approach to approximate and explain black-box DRL policies.
• To maximize open-loop performance and enhance closed-loop performance of the derived NLDT control rules.

Main Methods:

• Utilized a Nonlinear Decision Tree (NLDT) approach with a labeled state-action dataset from a pretrained DRL agent.
• Employed nonlinear optimization with evolutionary computation and a bilevel optimization procedure at each NLDT node.
• Proposed a reoptimization procedure for enhancing closed-loop performance and a postprocessing approach for NLDT simplification.

Main Results:

• The NLDT approach successfully generated hierarchical sets of simple, interpretable control rules (1-4 nonlinear terms per rule).
• Achieved closed-loop performance comparable to the original black-box DRL agent across various control problems.
• Demonstrated the potential to replace complex, non-interpretable DRL policies with simpler, understandable ones.

Conclusions:

• The proposed NLDT methodology offers a viable alternative to complex black-box DRL policies, providing interpretability without sacrificing performance.
• The findings are encouraging for applying NLDTs to more complex control tasks, enhancing trust and understanding in AI-driven control systems.