High Performance on Atari Games Using Perceptual Control Architecture Without Training

Abstract

AbstractDeep reinforcement learning (DRL) requires large samples and a long training time to operate optimally. Yet humans rarely require long periods of training to perform well on novel tasks, such as computer games, once they are provided with an accurate program of instructions. We used perceptual control theory (PCT) to construct a simple closed-loop model which requires no training samples and training time within a video game study using the Arcade Learning Environment (ALE). The model was programmed to parse inputs from the environment into hierarchically organised perceptual signals, and it computed a dynamic error signal by subtracting the incoming signal for each perceptual variable from a reference signal to drive output signals to reduce this error. We tested the same model across three different Atari games Breakout, Pong and Video Pinball to achieve performance at least as high as DRL paradigms, and close to good human performance. Our study shows that perceptual control models, based on simple assumptions, can perform well without learning. We conclude by specifying a parsimonious role of learning that may be more similar to psychological functioning.