Researchers use Overcooked to train AI to be a better gaming buddy

Two overcooked characters side by side.
(Image credit: Team 17)

DeepMind researchers have been using the chaotic cooking game Overcooked to teach AI to better collaborate with humans. MIT researchers have followed suit, gifting their AI the ability to distinguish between a diverse range of play-styles. What's amazing is that it's actually working—the humans involved actually preferred playing with the AI.

Have you ever been dropped into a game with strangers only to find their play-style totally upends your own? There's a reason we're better at gaming with people we know—they get us. As a team, you make a point of complementing each other's play-style so you can cover all bases, and win. 

Generally, an AI models' sole objective in gaming is to maximise its final score, screw the rest. That's why AI is great at single player competitive games, but not so much at collaborating. As we know, there's more to winning when it comes to co-op. 

DeepMind's researchers threw what they're calling 'Fictitious Co-Play (FCP)' AI into Overcooked sessions with both humans and 'novel agents,' or unfamiliar AI that was trained on separate algorithms.

Essentially it uses a much more diverse set to learn from. In their findings (PDF warning) they saw scores skyrocket thanks to this new training method, and the human partners even expressed "a strong subjective preference to partnering with FCP agents."

Virtual reality

(Image credit: Valve)

Best VR headset: which kit should you choose?
Best graphics card: you need serious GPU power for VR
Best gaming laptop: don't get tied to your desktop in VR

MIT researchers have been experementing with a similar method, inspired by the DeepMind findings. They took it a step further with their new 'Any-Play' model, broadening the criteria by asking the AI to correctly identify the play-style of its training partner, so it can better adapt for the good of the team. Rather than just maximising the score—which seems to be the goal of most of the boneheads I meet via online matchmaking.

Previous experiments that matched humans with AI teammates, such as one from Lincoln Laboratory researchers, made it clear that humans often see AI as a "confusing and unpredictable" teammate. These new methods, however, could see you one day be replaced by an AI, if your playstyle doesn't gel with your teammates'.

Katie Wickens
Hardware Writer

Screw sports, Katie would rather watch Intel, AMD and Nvidia go at it. Having been obsessed with computers and graphics for three long decades, she took Game Art and Design up to Masters level at uni, and has been demystifying tech and science—rather sarcastically—for three years since. She can be found admiring AI advancements, scrambling for scintillating Raspberry Pi projects, preaching cybersecurity awareness, sighing over semiconductors, and gawping at the latest GPU upgrades. She's been heading the PCG Steam Deck content hike, while waiting patiently for her chance to upload her consciousness into the cloud.