DeepMind’s new AI agents are learning to play Quake III Arena like humans

Deepmind AI agents are taught the strategies from scratch using three reinforcement learning ideas.

As the research in Artificial Intelligence (AI) furthers in the video gaming space, there is more scope for NPC or non-player characters to do better and work more efficiently. NPCs are video game characters that are controlled by the game's AI rather than by a gamer.

Researchers at DeepMind, a London-based artificial intelligence company, have taught their AI "agents" to achieve "human-level performance in Quake III Arena Capture the Flag," and these agents have the ability to team up with both AI players and human players.

According to a blog post by DeepMind on 3 July, the AI is taught the general strategies from scratch. It is focused on the Capture the Flag mode where the map changes in every match.

"Our agents must learn from scratch how to see, act, cooperate, and compete in unseen environments, all from a single reinforcement signal per match: whether their team won or not," DeepMind said in the post. It added that it is a challenging learning problem for which the solution is based on three general reinforcement learning ideas.

Quake III Arena Logo. Image: Quake III Arena YouTube

Quake III Arena Logo. Image: Quake III Arena YouTube

The first one is that instead of training the agents individually, they are put into groups, where they lean by playing with each other.

The second is that each agent is able to generate their own internal goals, such as capturing the flag.

The third and the final learning method is that the agents operate on two timescales, fast and slow, improving their ability to use memory and generate quick actions.

AI robots in gaming, are surprising all of us. We already know that at the Dota 2 world championship, The International, a top contender had quite a different looking team – a set of five AI robots.

"Our team of five neural networks, OpenAI Five, has started to defeat amateur human teams at Dota 2,” OpenAI mentioned in their most recent post on OpenAI blog.




Top Stories


also see

science