Accelerating Human-Agent Collaborative Reinforcement Learning

Abstract

In domains such as Human-Robot Collaboration artificial agents must be able to support mutual adaptation and learning. Towards this direction, we use a discrete Soft Actor-Critic agent on a real-time collaborative game with humans. We examine how different allocations of on-line and off-line gradient updates impact the game performance and the total training time. Our results suggest that early allocation of a high number of off-line g/u can accelerate learning while shortening training duration.