Human-level performance in 3D multiplayer games with population-based reinforcement learning

ABSTRACT

Multiplayer video games represent a significant frontier for AI research, requiring real-time, high-dimensional sensory processing, spatial navigation, and team-based coordination. We report an AI agent that achieves human-level performance in the popular 3D multiplayer first-person video game Quake III Arena Capture the Flag. The agent is trained exclusively from raw pixels and game points using a novel population-based multi-agent reinforcement learning approach. By concurrently training a diverse population of agents that continuously adapt their internal reward signals and hyperparameters, the system discovers complex macro-strategies, emergent teamwork, and high-level navigation skills, successfully defeating teams of highly skilled human players.