Probabilistic Verification Framework for RNN Policies in RL

Researchers propose RNN-ProVe, a probabilistic framework for verifying recurrent neural network policies in partially observable reinforcement learning. The method estimates the likelihood of undesired behaviors by using policy-driven sampling to approximate feasible hidden states, deriving statistical error bounds for high-confidence estimates. Experiments on single-agent and cooperative multi-agent tasks demonstrate its effectiveness. The work addresses challenges in verifying history-dependent policies that rely on latent hidden state dynamics, where existing tools often rely on restrictive assumptions or coarse over-approximations.