Spatio-temporal attention deep recurrent Q-network for POMDPs

Abstract

One of the long-standing challenges for reinforcement learning agents is to deal with noisy environments. Although progress has been made in producing an agent capable of optimizing its environment in fully observable conditions, partial observability still remains a difficult task. In this paper, a novel model is proposed which inspired by human perception, utilizes two fundamental machine learning concepts, attention and memory, to better confront a noisy environment.