Grandmaster level in StarCraft II using multi-agent reinforcement learning.

Area of Science:

• Artificial Intelligence
• Computational Game Theory
• Multi-Agent Systems

Background:

• StarCraft presents complex multi-agent challenges relevant to real-world applications.
• Previous AI agents in StarCraft have relied on game simplification or superhuman abilities.
• No prior AI has matched the overall skill of top human StarCraft players.

Purpose of the Study:

• To develop an AI agent capable of competing at a top human level in the complex real-time strategy game StarCraft II.
• To utilize general-purpose learning methods applicable to other challenging domains.

Main Methods:

• Employed a multi-agent reinforcement learning algorithm.
• Utilized a diverse league of deep neural networks with adaptive strategies and counter-strategies.
• Trained on data from both human and agent games in StarCraft II.

Main Results:

• The AlphaStar agent achieved Grandmaster level across all three StarCraft II races.
• AlphaStar's performance surpassed 99.8% of ranked human players.
• Demonstrated a significant advancement in AI capabilities for complex strategy games.

Conclusions:

• General-purpose learning methods, particularly multi-agent reinforcement learning, can achieve human-level performance in complex strategic environments like StarCraft II.
• AlphaStar represents a major milestone in artificial intelligence research for real-time strategy games.
• The approach is extensible to other domains requiring complex decision-making and coordination.