DeepMind AI agents learn to beat humans at Quake 3

July 6, 2018 — by Manish "Trigger-Happy" Rajesh0

Bots have surpassed humans, in Quake 3 at least, on a stripped-down version of Quake 3 Arena’s Capture the Flag mode.

Quake 3

Similar to the OpenAI Dota 2 bots that are looking to beat pro teams at The International, Google’s DeepMind has used its reinforced learning to teach AI the multiplayer CtF mode in Quake 3 Arena. The bots have supposedly played around 450,000 games against themselves, each lasting five minutes on a procedurally generated map. That’s 37,500 hours of practice. Way more than any human can put in.

According to the Verge, unlike OpenAI, DeepMind did not have access to raw numerical data on Quake 3, but used visual on-screen inputs, the same way an actual human player might. The DeepMind agents also were given no instructions. They just played against one another till it was determined that they could most probably win.

It’s crazy. The bots not only worked out strategies and tactics, but used techniques like ganking, and camping among others.

DeepMind hosted tournaments with player teams of humans, two player teams of robots and a combination of both. The bot only team had a massive 74 per cent win probability vs a 52 per cent for skilled human players and 43 per cent for average players.

When the number of bots on a team was increased to four, the win per cent probability dropped to 64, but that’s still higher than the avg for human players.



Manish "Trigger-Happy" Rajesh

If he's not gaming, he's... no wait he's always gaming.

Leave a Reply