In its bid to show how smart AI can be, DeepMind introduces its new agent called MuZero that can master simple to complex games without even being told the rules.
The recent project isn’t DeepMind’s first attempt at creating a game-playing AI bot. Previously, it released programs AlphaGo, AlphaGo Zero, and AlphaZero. However, MuZero stands out.
Unlike the programs that came before it, MuZero isn’t provided with a rulebook prior to playing games. The final model learned the roles from experimenting on its own, without human data or interaction. Most AI agents would use the rules to find best-case scenarios – something that DeepMind’s project couldn’t do.
To make up for its lack of information, MuZero factors in the game environment to understand the concepts better. After it was fed millions of games, the AI agent learned the rules, the value of the positions, and the general policies to reevaluate its past actions. Simply put, DeepMind’s model banks on its ability to learn from its own mistakes and redos the games until it polishes its knowledge.
According to DeepMind, MuZero shows AI’s ability to intelligently interact with the real word, learn how to understand it, and respond without always being told what to do.