DeepMind AI agents are getting pretty good at Quake 3 Arena

A new test using DeepMind AI has shown that these agents are getting pretty good at Quake 3 Arena.


Google’s DeepMind AI has always been a particularly interesting thing. While we’ve already seen the DeepMind AI pull off some intriguing things, the latest tests of the agents test their ability to work together and learn within an environment like Capture the Flag. The results of the test were not really all that surprising, but still astounding, nonetheless.

According to the report, the DeepMind AI agents were able to comprehensively work and learn within a series of randomly generated Quake 3 Arena maps. Each map followed standard Capture The Flag mechanics, and required the agents to work together to win. The version of Quake 3 Arena that the test utilized was customized, however all of the basic game mechanics still remained the same. The idea behind the test was to see how well the agents could work in a team of other agents, as well as how they would work when playing alongside human players.

How DeepMind AI represent the game world around them.
How DeepMind AI represent the game world around them.

What’s most interesting about the agents used in these tests, is that they must learn everything from scratch. They aren’t programmed to know how to aim, or where everything is. They have to learn to act, cooperative, and even compete in environments that they do not know. On top of this, each agent is also set to learn its own internal reward signal. So, for some it might be capturing the flag, while others might go for a different internal reward goal, like defending a teammate or something of that nature.

Based on the records from the report, the testers ran a tournament including 40 human players. In the tournament, human players and agents were randomly matched into games as both opponents and as teammates. Surveys taken after the tournament showed that the agents were able to exceed the win-rate of the human players, and many of the human participants stated that they felt more collaboration between the agents than their human participants.

The entire report on the latest tests can be read on the DeepMind website. It’s a very long, but interesting read, especially if you’re intrigued by the future of artificial intelligence at all. Make sure you check out the detailed results and be sure to follow Shacknews on Twitter for more coverage of news, features, and strategy content. You can join the conversation by heading over to the original Chatty thread here, or by making a comment below.

Guides Editor

Joshua holds a Bachelor of Fine Arts in Creative Writing and has been exploring the world of video games for as long as he can remember. He enjoys everything from large-scale RPGs to small, bite-size indie gems and everything in between.

From The Chatty
Hello, Meet Lola