Beyond being known as a search engine, Google is also responsible for creating DeepMind, one of the world’s most advance AI that defeated both Korean and Chinese players in a one-on-one game of the ancient Chinese game of Go. Now it appears that Google is training DeepMind how to play Quake III Arena.
To be precise, DeepMind is training multiple AI agents in its system to play the game, and the reason behind this boils down to one word: teamwork. In-game, the AI agents are being trained to work as teammates through multiple rounds of the game’s “Capture the Flag” mode. At the time of writing, Google says that the AI – named For the Win (FTW) – had clocked in close to 450,000 games learning how to see, act, cooperate, and compete in constantly changing environments.
The acid test came when Google then held a tournament with 40 human players. These players were matched with AI agents at random, both as opponents and as teammates. Over the course of the tournament, it was observed that the teams consisting entirely of FTW agents performed the best.
On that note, Google says that its goal for doing this has less to do with winning esports tournaments, and more to do with seeing if an AI system is capable of learning to work as a team when a game mode requires it to cooperate and interact with other agents, be it AI or human.
This isn’t the first time that DeepMind has been subjected to learning how to play a game. Back in 2016, the company collaborated with video game developer Blizzard to train its AI on playing Starcraft II. Unlike Quake III Arena, DeepMind’s focus on the RTS is more about getting its AI to defeat a human player in a one-on-one game.