Decision-Aware Actor-Critic with Function Approximation and Theoretical Guarantees (original) (raw)

2023, arXiv (Cornell University)

Actor-critic (AC) methods are widely used in reinforcement learning (RL), and benefit from the flexibility of using any policy gradient method as the actor and value-based method as the critic. The critic is usually trained by minimizing the TD error, an objective that is potentially decorrelated with the true goal of achieving a high reward with the actor. We address this mismatch by designing a joint objective for training the actor and critic in a decision-aware fashion. We use the proposed objective to design a generic, AC algorithm that can easily handle any function approximation. We explicitly characterize the conditions under which the resulting algorithm guarantees monotonic policy improvement, regardless of the choice of the policy and critic parameterization. Instantiating the generic algorithm results in an actor that involves maximizing a sequence of surrogate functions (similar to TRPO, PPO), and a critic that involves minimizing a closely connected objective. Using simple bandit examples, we provably establish the benefit of the proposed critic objective over the standard squared error. Finally, we empirically demonstrate the benefit of our decision-aware actor-critic framework on simple RL problems. 1 Introduction Reinforcement learning (RL) is a framework for solving problems involving sequential decisionmaking under uncertainty, and has found applications in games [38, 50], robot manipulation tasks [55, 64] and clinical trials [45]. RL algorithms aim to learn a policy that maximizes the long-term return by interacting with the environment. Policy gradient (PG) methods [59, 54, 29, 25, 47] are an important class of algorithms that can easily handle function approximation and structured state-action spaces, making them widely used in practice. PG methods assume a differentiable parameterization of the policy and directly optimize the return with respect to the policy parameters. Typically, a policy's return is estimated by using Monte-Carlo samples obtained via environment interactions [59]. Since the environment is stochastic, this approach results in high variance in the estimated return, leading to higher sample-complexity (number of environment interactions required to learn a good policy). Actor-critic (AC) methods [29, 43, 5] alleviate this issue by using value-based approaches [52, 58] in conjunction with PG methods, and have been empirically successful [20, 23]. In AC algorithms, a value-based method ("critic") is used to approximate a policy's estimated value, and a PG method ("actor") uses this estimate to improve the policy towards obtaining higher returns. Though AC methods have the flexibility of using any method to independently train the actor and critic, it is unclear how to train the two components jointly in order to learn good policies. For example, the critic is typically trained via temporal difference (TD) learning and its objective is to minimize the value estimation error across all states and actions. For large real-world Markov decision processes (MDPs), it is intractable to estimate the values across all states and actions, and 37th Conference on Neural Information Processing Systems (NeurIPS 2023).