Human-level Atari 200x faster
The task of building general agents that perform well over a wide range of tasks has been an importantgoal in reinforcement learning since its inception. The problem has been subject of research of alarge body of work, with performance frequently measured by observing scores over the wide rangeof environments contained in the Atari 57 benchmark. Agent57 was the first agent to surpass thehuman benchmark on all 57 games, but this came at the cost of poor data-efficiency, requiring nearly 80billion frames of experience to achieve. Taking Agent57 as a starting point, we employ a diverse set ofstrategies to achieve a 200-fold reduction of experience needed to outperform the human baseline. Weinvestigate a range of instabilities and bottlenecks we encountered while reducing the data regime, andpropose effective solutions to build a more robust and efficient agent. We also demonstrate competitiveperformance with high-performing methods such as Muesli and MuZero. The four key components toour approach are (1) an approximate trust region method which enables stable bootstrapping from theonline network, (2) a normalisation scheme for the loss and priorities which improves robustness whenlearning a set of value functions with a wide range of scales, (3) an improved architecture employingtechniques from NFNets in order to leverage deeper networks without the need for normalization layers,and (4) a policy distillation method which serves to smooth out the instantaneous greedy policy overtime.
PDF Abstract
Arcade Learning Environment
