Enabling Realtime Reinforcement Learning at Scale with Staggered Asynchronous Inference

TL;DR

This paper analyzes the limits of real-time reinforcement learning, proposing asynchronous inference algorithms that enable larger models and consistent action timing in fast-changing environments like games.

Contribution

It introduces staggered asynchronous inference algorithms and provides theoretical analysis showing their effectiveness in real-time RL with large models.

Findings

Asynchronous inference reduces regret in real-time RL environments.

Models with longer inference times can be used effectively with the proposed methods.

Scalability to larger models is achieved without sacrificing reaction time.

Abstract

Realtime environments change even as agents perform action inference and learning, thus requiring high interaction frequencies to effectively minimize regret. However, recent advances in machine learning involve larger neural networks with longer inference times, raising questions about their applicability in realtime systems where reaction time is crucial. We present an analysis of lower bounds on regret in realtime reinforcement learning (RL) environments to show that minimizing long-term regret is generally impossible within the typical sequential interaction and learning paradigm, but often becomes possible when sufficient asynchronous compute is available. We propose novel algorithms for staggering asynchronous inference processes to ensure that actions are taken at consistent time intervals, and demonstrate that use of models with high action inference times is only constrained by the environment's effective stochasticity over the inference horizon, and not by action frequency. Our analysis shows that the number of inference processes needed scales linearly with increasing inference times while enabling use of models that are multiple orders of magnitude larger than existing approaches when learning from a realtime simulation of Game Boy games such as Pok\'emon and Tetris.