Researchers from DeepMind, a division of Alphabet in the field of studying AI, reported about a new development designed to teach AI better to play video games. This time, the experts modified Quake III Arena and its "Capture Flag" mode and forced AIs to learn this game.
Mastering the strategy, tactical understanding, and team play involved in multiplayer video games represents a critical challenge for AI research. Now, through new developments in reinforcement learning, our agents have achieved human-level performance in Quake III Arena Capture the Flag, a complex multi-agent environment and one of the canonical 3D first-person multiplayer games. These agents demonstrate the ability to team up with both artificial agents and human players.
The authors chose "Flag Capture" for AI agents to learn the mechanics of the game in a procedurally-generated world. Agents played both alone and gathering in teams, including with people. In addition, AI has learned to use tactics such as protecting the base, waiting for the enemy and following the partner:
The developers used the training method with reinforcement, and the AI did not receive any additional information, except for the picture on the screen. The team of agents trained with each match, receiving a positive response when winning. At the same time, each of them had its own internal reward. AI is based on a pair of recurrent neural networks, fast and slow, each of which studies the transition from scoring points to an internal reward.
According to the results of the study, the authors found that AI agents not only won more often than people, but also were more united. According to the received data, the AI coefficient Elo, responsible for the chance of winning, is higher than the human:
The authors claim that in the future they will develop technologies of simultaneous training with the reinforcement of several AI agents, and will also pay more attention to uniting agents and people in teams for greater efficiency.