A Unified Causal-Origin Taxonomy of Distributional Shifts in Reinforcement Learning

This work presents a unified causal-origin taxonomy for understanding distributional shifts in reinforcement learning (RL), linking In-Distribution (ID) and Out-of-Distribution (OOD) generalization with non-stationary environments. It reformulates distributional shifts using a Partially Observable Markov Decision Process (POMDP), identifying internal and external sources of shifts, and introduces an evaluation framework for assessing performance impacts and recovery metrics. This framework provides practitioners with a structured approach to analyze and enhance the robustness of RL systems under varying operational conditions.