AlphaStar Unplugged: Large-Scale Offline Reinforcement Learning

TL;DR

This paper introduces AlphaStar Unplugged, a large-scale offline RL benchmark for StarCraft II, leveraging a massive human gameplay dataset to develop and evaluate offline RL agents in a complex, real-time strategy environment.

Contribution

It establishes a new offline RL benchmark for StarCraft II, providing datasets, tools, and evaluation protocols, and demonstrates improved agents using only offline data.

Findings

Achieved 90% win rate against previous behavior cloning agents.

Established a standardized API and evaluation protocol for offline StarCraft II RL.

Presented baseline agents including behavior cloning, offline actor-critic, and MuZero variants.

Abstract

StarCraft II is one of the most challenging simulated reinforcement learning environments; it is partially observable, stochastic, multi-agent, and mastering StarCraft II requires strategic planning over long time horizons with real-time low-level execution. It also has an active professional competitive scene. StarCraft II is uniquely suited for advancing offline RL algorithms, both because of its challenging nature and because Blizzard has released a massive dataset of millions of StarCraft II games played by human players. This paper leverages that and establishes a benchmark, called AlphaStar Unplugged, introducing unprecedented challenges for offline reinforcement learning. We define a dataset (a subset of Blizzard's release), tools standardizing an API for machine learning methods, and an evaluation protocol. We also present baseline agents, including behavior cloning, offline variants of actor-critic and MuZero. We improve the state of the art of agents using only offline data, and we achieve 90% win rate against previously published AlphaStar behavior cloning agent.