DeepMind researchers have taught artificially intelligent gamers to play a popular 3D multiplayer first-person video game with human-like skills - a previously insurmountable task. The reinforcement learning-trained AI agents demonstrate an uncanny ability to develop and use independently learned high-level strategies to compete and cooperate in the game environment. Reinforcement learning (RL), a method used to train artificially intelligent agents, has shown success in producing artificially intelligent players that are able to navigate increasingly complex single-player environments. These agents can also achieve superhuman mastery in competitive two-player turn-based games, like chess and Go. However, the ability to play multiplayer games, particularly those that involve teamwork and interaction between multiple independent players, has eluded the capabilities of AI systems to date. Here, Max Jaderberg and colleagues present a RL-trained AI agent that can achieve human-level performance in the seminal multiplayer 3D first-person video game, Quake III Arena Capture the Flag. In contrast to previous studies, where AI agents are supplied with "knowledge" about the game environment or status of other players, Jaderberg et al.'s RL approach ensured that each agent learned independently from its own experience using only what it could "see" (pixels) and the game's score. Pitted against one another, a population of AI agents learned to play the game over thousands of matches in randomly generated environments. According to the authors, over time, the agents independently developed surprisingly high-level strategies, not unlike those used by skilled human players. What's more, in games with human players, the agents outperformed human opponents, even when the agents' reaction times were slowed down to human levels. Furthermore, the agents were also able to form and cooperate in ad hoc computer and human teams.
###
Journal
Science