ORCAID extracts interpretable, rule-based policies from deep reinforcement learning agents in mixed continuous-discrete environments with continuous actions. It trains oblique decision trees that partition the state space with hyperplanes and fit local linear models in the leaves. Its three-stage split search combines efficient random initialization, local refinement, and backward elimination, followed by adjacent-leaf merging to produce concise rules. The authors report strong retained performance across multiple RL environments and claim that the distilled policy can sometimes improve the original deep RL policy.
No heat snapshots are available in the last 24 hours.