SPOT: New Framework for Interpreting Deep Reinforcement Learning Policies

A recent publication on arXiv (2608.09967) presents SPOT (Sampling Policy Observation Tree), a framework that is model-agnostic and sampling-based for the interpretation of deep reinforcement learning (DRL) policies. By sampling actions and recursively simulating successor states, this framework builds an interpretable finite-horizon tree, utilizing the policy and an environment simulator. This tree empirically illustrates the policy's action preferences and their possible future developments. The authors provide formal assurances that SPOT can asymptotically identify the policy's most likely action and analyze its disagreement behavior in high-entropy scenarios. Demonstrated within the SUMO-RL traffic-signal control domain, the tree representation aids in evaluating policy behavior, tackling the challenge of deciphering the often opaque decision-making processes of DRL agents.