Meta attention for Off-Policy Actor-Critic.

Abstract

Off-Policy Actor-Critic methods can effectively exploit past experiences and thus they have achieved great success in various reinforcement learning tasks. In many image-based and multi-agent tasks, attention mechanism has been employed in Actor-Critic methods to improve their sampling efficiency. In this paper, we propose a meta attention method for state-based reinforcement learning tasks, which combines attention mechanism and meta-learning based on the Off-Policy Actor-Critic framework. Unlike previous attention-based work, our meta attention method introduces attention in the Actor and the Critic of the typical Actor-Critic framework, rather than in multiple pixels of an image or multiple information sources in specific image-based control tasks or multi-agent systems. In contrast to existing meta-learning methods, the proposed meta-attention approach is able to function in both the gradient-based training phase and the agent's decision-making process. The experimental results demonstrate the superiority of our meta-attention method in various continuous control tasks, which are based on the Off-Policy Actor-Critic methods including DDPG and TD3.