Artificial intelligence “agents” began behaving like humans after teaching themselves how to team up and outwit people in a bid to win a popular multi-player video game, researchers have revealed.
The computer system, created by Google-owned DeepMind, was trained to play Quake III Arena Capture the Flag, where two teams of individual players navigate a maze with the goal of capturing the opponent team’s flag.
Humans have spent years playing against computers in two-player games. But DeepMind’s AI appeared to pick up typically human psychological tactics in order to win games – rather than memorizing the game map and possible outcomes from a move. After teaching themselves how to play, agents began to follow team-mates and camp out in opponents’ base camp to try to send them back to the beginning.
In the game, players can tag opponents, sending them back to the start. The team that captures the flag the most in five minutes takes victory. DeepMind ran a tournament including 40 human players, both as opponents and team-mates, which allowed the AI to pick up tricks such as positioning itself in an opponent’s base to send them back to the beginning, and following team-mates to provide backup.
A population of agents was trained while playing thousands of games concurrently in parallel. The agents went on to become stronger and exceeded the win rate of human players.
A survey among participants found that the computer system was more collaborative than its human -counterparts.
Researchers used “multi-agent training”, a technique that they hope will advance the development of artificial intelligence so that it can begin to team up with humans for other tasks.
The company, spun out of Cambridge University by three computer science students, famously created a computer program that defeated a Go champion in 2016.
DeepMind was bought by Google in 2014. It continues to operate in London, where the majority of staff work.
The company is focused on creating artificial intelligence that could benefit the health sector, but has applied its research to Google’s day-to-day running, including reducing the impact its massive data servers are having on the environment by improving their cooling functions.