Go to ICLR 2025 Conference homepageDRIVE: Distributional Model-Based Reinforcement Learning via Variational Inference
Distributional reinforcement learning (RL) provides a natural framework for estimating the distribution of returns rather than a single expected value. However, the control aspect of distributional RL has not been as thoroughly explored as the evaluation part, typically relying on the greedy selection rule with respect to either the expected value, akin to standard approaches, or risk-sensitive measures derived from the return distribution. On the other hand, casting RL as a probabilistic inference problem allows for flexible control solutions utilizing a toolbox of approximate inference techniques; however, its connection to distributional RL remains underexplored. In this paper, we bridge this gap by proposing a variational approach for efficient policy search. Our method leverages the log-likelihood of optimality as a learning proxy, decoupling it from traditional value functions. This learning proxy incorporates aleatoric uncertainty of the return distribution, enabling risk-aware decision-making. We provide a theoretical analysis of our framework, detailing the conditions for convergence. Empirical results on vision-based tasks in DMControl Suite demonstrate the effectiveness of our approach compared to various algorithms, as well as its ability to balance exploration and exploitation at different training stages.