Robust and Risk-Sensitive Reinforcement Learning: A Systematic Empirical Evaluation of the Deployment Gap

This research presents a comprehensive survey and systematic empirical evaluation of state-of-the-art paradigms in robust and risk-sensitive Reinforcement Learning (RL). The study aims to address the “deployment gap” — the discrepancy between controlled training environments and the stochastic, often mismatched, dynamics of real-world systems.

The thesis establishes a unified evaluation infrastructure, defining standardized metrics for performance, robustness, and computational practicality across both model-based and model-free RL approaches. The methodology employs rigorous sensitivity analysis regarding environmental stochasticity, model mismatch, and hyperparameter variation. Furthermore, ablation studies are conducted to isolate the marginal contributions of key algorithmic components, including risk-sensitive objectives, critics, and model uncertainty estimation. By providing a granular failure analysis for algorithm-environment pairs across diverse discrete and continuous control benchmarks, this work offers a diagnostic framework for assessing the reliability and limits of current safe RL methodologies.

M.Sc. student under the supervision of Dr. Sarah keren.