Prediction-Based Markov Violation Scores for Detecting Non-Markovian Observations in Reinforcement Learning

arXiv:2603.27389v2 Announce Type: replace-cross Abstract: Reinforcement learning algorithms assume that observations satisfy the Markov property, yet real-world sensors frequently violate this assumption through correlated noise, latency, or partial observability. Standard performance metrics...