Distributional Model Equivalence for Risk-Sensitive Reinforcement Learning
This work addresses the challenge of optimal planning under risk measures for reinforcement learning practitioners, offering a novel theoretical and practical framework.
The paper tackles the problem of learning models for risk-sensitive reinforcement learning by showing that existing risk-neutral methods are insufficient and introducing new distributional model equivalence notions, with experiments demonstrating practical applicability in tabular and large-scale settings.
We consider the problem of learning models for risk-sensitive reinforcement learning. We theoretically demonstrate that proper value equivalence, a method of learning models which can be used to plan optimally in the risk-neutral setting, is not sufficient to plan optimally in the risk-sensitive setting. We leverage distributional reinforcement learning to introduce two new notions of model equivalence, one which is general and can be used to plan for any risk measure, but is intractable; and a practical variation which allows one to choose which risk measures they may plan optimally for. We demonstrate how our framework can be used to augment any model-free risk-sensitive algorithm, and provide both tabular and large-scale experiments to demonstrate its ability.