Risk-Sensitive Variational Model-Based Policy Optimization

Abstract

RL-as-inference casts reinforcement learning (RL) as Bayesian inference in a probabilistic graphical model. While this framework allows efficient variational approximations it is known that model-based RL-as-inference learns optimistic dynamics and risk-seeking policies that can exhibit catastrophic behavior. By exploiting connections between the variational objective and a well-known risk-sensitive utility function we adaptively adjust policy risk based on the environment dynamics. Our method, $\beta$-VMBPO, extends the variational model-based policy optimization (VMBPO) algorithm to perform dual descent on risk parameter $\beta$. We provide a thorough theoretical analysis that fills gaps in the theory of model-based RL-as-inference by establishing a generalization of policy improvement, value iteration, and guarantees on policy determinism. Our experiments demonstrate that this risk-sensitive approach yields improvements in simple tabular and complex continuous tasks, such as the DeepMind Control Suite.