DeepMind develops AI that allows you to play various games like humans, and also supports multiplayer
Open-Ended Learning Leads to Generally Capable Agents | DeepMind
https://deepmind.com/research/publications/open-ended-learning-leads-to-generally-capable-agents
Generally capable agents emerge from open-ended play | DeepMind
https://deepmind.com/blog/article/generally-capable-agents-emerge-from-open-ended-play
The Go AI 'AlphaGo' developed by DeepMind continues to make progress even after defeating the world's strongest Go player, and in 2017, ' AlphaZero ' that supports not only Go but also shogi and chess games will appear. Attempts have also been made to use this AlphaZero to 'create new chess rules by learning anomalous rules.'
AI that defeated humans and changed the way chess is done is used to 'open up new possibilities for chess' --GIGAZINE
However, according to DeepMind, although AlphaZero supports multiple games such as Go, Shogi, and chess, each game requires different learning. DeepMind said, 'We sought a way to overcome the limitations of AlphaZero and develop highly adaptable AI agents.' 'Newly developed AI is more than specific to a specific task. This is a significant step towards developing an AI that can respond quickly to ever-changing situations. ”He announced that he has developed an AI that can handle all kinds of games. Did.
The AI developed this time was designed to operate characters from a first-person perspective, and learned how to achieve relatively simple goals in multiplayer games such as 'flag-taking games,' 'hide-and-seek,' and 'color matching games.' The goals given to AI are 'carry a yellow cube to a defined area' in a flag-taking game, 'out of sight of an enemy player' in hide-and-seek, and 'close objects of the same color' in a color matching game. It depends on the game you learn. In addition, the stage of the game can be changed randomly by the program, and AI also learned how to grasp the map of the stage from the color information around the character that it operates.
As a result of learning AI 200 billion times by the above method, it seems that AI is not 'highly optimized behavior for a specific task' but 'judgment by empirical rules like human beings' It is said that it has come to show 'good behavior'. In addition, AI has also learned movements unique to multiplayer, such as 'using objects in the map to block the view of other players.'
In the following movie, you can see how the AI developed this time actually plays the game.
In 'Kakurenbo', the blue player aims to bring the red player into view, and the red player aims to escape from the blue player's view.
In the movie, the image when the map is viewed from above and the blue player viewpoint image are displayed. Immediately after the game starts, the blue player keeps the red player in sight ...
After a while from the start of the game, the red player escapes from the blue player's field of view.
After that, the red player managed to escape to the back of the map.
In the flag-taking game, both players aim to bring two cubes onto a base that matches their color.
Players can use the beam to defeat enemy players. When the game actually starts, a fierce beam exchange immediately after the start.
In the end, the blue player succeeded in bringing the cube to his side.
In the next game, both players aim to touch the purple quadrangular pyramid.
There are many ways to touch the purple quadrangular pyramid. For example, making a slope with a large plate-shaped object ...
You can see how AI achieves its purpose in various ways, such as hitting an object on a purple quadrangular pyramid and dropping it in a low position. The AI developed this time is not specialized for a single game, but is proficient in all of the above games.
DeepMind invites those who are interested in research to participate in the research.
Related Posts:
