Multi-task reinforcement learning with associative memory models considering the multiple distributions of MDPs

Abstract

Multi-task reinforcement learning is one of the promising approaches in reinforcement learning problems. While the formulation of the multi-task reinforcement learning problem have been established in a previous study, only a single distribution of the tasks has been considered. However, we assume that the formulation can hardly be applied to real-world problems. This paper presents a method of expanding the formulation to a more general problem by considering multiple distributions of tasks. In addition, we propose an agent model with associative memory models, then apply it to an expanded multi-task reinforcement learning problem.