AI bots have been a long-standing part of multiplayer video games, tagging in when there’s no local or online players available. However, for the most part, they suck, often making the entire experience more frustrating than it should be. Researchers at MIT are set to change that with a co-op game AI that’s been trained to be a perfect companion.
How MIT trained a co-op game AI
In a collaboration with Google DeepMind, MIT researchers have worked to create a “Fictitious Co-Play (FCP)” AI. Designed as a co-op game AI buddy, the technology could one day be used to make video game bots less useless.
Detailed in their study “Collaborating with Humans without Human Data”, the research requires heavy training of an AI to play games. At first, this would require the AI to learn how to “self-play” before being able to assist humans.
After many hours of training for different play styles, the FCP AI was turned into the perfect co-op partner. While previous generations of machine learned AI fostered frustration, experiments with the new co-op game AI were positive.
After extensive testing with human players, many preferred having the AI player to the alternative. The study claims that players had a “strong subjective preference to partnering with FCP agents."
When will this technology be used in games?
At the time of writing, DeepMind and MIT’s is trained only for Overcooked. However, it’s neural net should be able to be multipurposed for other games after training.
If successfully trained, developers could use the AI as a baseline for future co-operative video games. For example, the next Overcooked could have bots that make you less bad at the game. Now wouldn’t that be a miracle?
It’ll be interesting to see just how far game AI can be pushed now that true AIs are on the way.