Researchers used a team of AI agents with reinforcement learning to beat humans in capture the flag mode in Quake III. (Image credit: DeepMind)


Alphabet’s DeepMind AI has already beaten the best at Go, Shogi, and chess, and now Quake III can be added to its list of gaming achievements. The difference this time around, is that instead of just one player playing against another, it was team VS team in the popular first-person shooter’s capture the flag mode. In the game, teams are pitted against each other in an effort to capture all of the opposing team’s flags (or most), and returning them to their base, all the while protecting their own flags in the process- thus winning the round.


In a recently released paper, DeepMind researchers explain how they were able to apply the mechanics of teamwork for “AI agents” or players, which can be on either team- synthetic or human. In essence, the researchers trained those agents by making them play 450,000 rounds of capture the flag, a timeframe that would take humans roughly four years, and which they did in only a few weeks using reinforcement learning.



The AI agents begin their training by employing random movements around an arena map, and after playing repeatedly, the agents start to strategize and form techniques that make those movements efficient on that particular map. They then take that reinforcement learning process and apply it in every arena, using teamwork to accomplish their goals.


The researchers ran a Quake III tournament, pitting agents against humans, and found that the AI team successfully competed against their human counterparts even when their reaction times were slowed to match their opponents. The AI win rate was much higher than the humans using their reinforcement learning technique.  Interesting enough, according to the paper- when integrated with a human team, the agents adapt to their arbitrary skills and still have a higher win rate than an all-human team.  


Have a story tip? Message me at: cabe(at)element14(dot)com