Out-of-Distribution Detection in RL

The paper, Rethinking Out-of-Distribution (OOD) Detection for Reinforcement Learning, presents new benchmarks for evaluating RL agents’ ability to detect novel scenarios.

• Proposes new terminology for OOD detection in RL.
• Introduces benchmark scenarios with temporal autocorrelation anomalies.
• Develops DEXTER, a method combining time series data analysis with anomaly detection.

This paper challenges existing OOD detection techniques and introduces DEXTER as a potential solution, promising more robust response mechanisms for RL agents in unanticipated environments.