This paper proposes an algorithm that aims to improve generalization for reinforcement learning agents by removing overfitting to confounding features. Our approach consists of a max-min game theoretic objective. A generator transfers the style of observation during reinforcement learning. An additional goal of the generator is to perturb the observation, which maximizes the agent's probability of taking a different action. In contrast, a policy network updates its parameters to minimize the effect of such perturbations, thus staying robust while maximizing the expected future reward. Based on this setup, we propose a practical deep reinforcement learning algorithm, Adversarial Robust Policy Optimization (ARPO), to find a robust policy that generalizes to unseen environments. We evaluate our approach on Procgen and Distracting Control Suite for generalization and sample efficiency. Empirically, ARPO shows improved performance compared to a few baseline algorithms, including data augmentation.

Robust policy optimization acknowledges that risk-aversion plays a vital role in real-world decision-making. When faced with uncertainty about the effects of actions, the policy that maximizes expected utility over the unknown parameters of the system may also carry with it a risk of intolerably poor performance. One might prefer to accept lower utility in expectation in order to avoid, or reduce the likelihood of, unacceptable levels of utility under harmful parameter realizations. In this paper, we take a Bayesian approach to parameter uncertainty, but unlike other methods avoid making any distributional assumptions about the form of this uncertainty. Instead we focus on identifying optimization objectives for which solutions can be efficiently approximated.

The policy gradient method enjoys the simplicity of the objective where the agent optimizes the cumulative reward directly. Moreover, in the continuous action domain, parameterized distribution of action distribution allows easy control of exploration, resulting from the variance of the representing distribution. Entropy can play an essential role in policy optimization by selecting the stochastic policy, which eventually helps better explore the environment in reinforcement learning (RL). However, the stochasticity often reduces as the training progresses; thus, the policy becomes less exploratory. Additionally, certain parametric distributions might only work for some environments and require extensive hyperparameter tuning. This paper aims to mitigate these issues. In particular, we propose an algorithm called Robust Policy Optimization (RPO), which leverages a perturbed distribution. We hypothesize that our method encourages high-entropy actions and provides a way to represent the action space better. We further provide empirical evidence to verify our hypothesis. We evaluated our methods on various continuous control tasks from DeepMind Control, OpenAI Gym, Pybullet, and IsaacGym. We observed that in many settings, RPO increases the policy entropy early in training and then maintains a certain level of entropy throughout the training period. Eventually, our agent RPO shows consistently improved performance compared to PPO and other techniques: entropy regularization, different distributions, and data augmentation. Furthermore, in several settings, our method stays robust in performance, while other baseline mechanisms fail to improve and even worsen the performance. The policy gradient method directly optimizes the expected return, making them easy to understand, stable to train, and effective. One advantage of this approach is that the exploration can be controlled by representing the action space with parameterized distribution (e.g., Gaussian). Due to this heavy dependency on action distribution, it is important that the parameterized distribution adequately represent the underlying action of the task.

Robust policy optimization acknowledges that risk-aversion plays a vital role in real-world decision-making. When faced with uncertainty about the effects of actions, the policy that maximizes expected utility over the unknown parameters of the system may also carry with it a risk of intolerably poor performance. One might prefer to accept lower utility in expectation in order to avoid, or reduce the likelihood of, unacceptable levels of utility under harmful parameter realizations. In this paper, we take a Bayesian approach to parameter uncertainty, but unlike other methods avoid making any distributional assumptions about the form of this uncertainty. Instead we focus on identifying optimization objectives for which solutions can be efficiently approximated.