In the classical Reinforcement Learning (RL) setting, one aims to find a policy that maximizes its expected return. This objective may be inappropriate in safety-critical domains such as healthcare or autonomous driving, where intrinsic uncertainties due to stochastic policies and environment variability may lead to catastrophic failures. This can be addressed by using the Conditional-Value-at-Risk (CVaR) objective to instill risk-aversion in learned policies. In this paper, we propose Adversarial Cvar Reinforcement Learning (ACReL), a novel adversarial meta-algorithm to optimize the CVaR objective in RL. ACReL is based on a max-min between a policy player and a learned adversary that perturbs the policy player's state transitions given a finite budget. We prove that, the closer the players are to the game's equilibrium point, the closer the learned policy is to the CVaR-optimal one with a risk tolerance explicitly related to the adversary's budget. We provide a gradient-based training procedure to solve the proposed game by formulating it as a Stackelberg game, enabling the use of deep RL architectures and training algorithms. Empirical experiments show that ACReL matches a CVaR RL state-of-the-art baseline for retrieving CVaR optimal policies, while also benefiting from theoretical guarantees.
View on arXiv