MuZero learns and plans not unlike a human being.
Remember how, at the end of WarGames, they have to teach the big computer about mutually-assured destruction by pitting it against itself in infinite matches of Tic-Tac-Toe? What if the computer could infer such a thing from the get-go instead of needing to be taught? Well, that movie probably would’ve been shorter. But the point is that an AI that could infer the rules of a game from the beginning could make much more educated decisions about how to play it. That’s what so impressive about DeepMind’s recent development.
DeepMind, a subsidiary of Alphabet, has been tinkering with game-playing artificial intelligences for a few years now. Its first few offerings, Alpha Go, Alpha Go Zero, and AlphaZero, could play board games like Go, Chess, and Shogi at expert levels. Thing is, all three of these AIs needed to be explicitly taught the rules of the games to build up their algorithms, which took a lot of trial and error. DeepMind’s newest AI, MuZero, is able to play all of those games, as well as simple Atari games, without needing to be told the rules at all. By just presenting the AI with the game, it’s able to infer its rules and quickly master them. This is a concept AI engineers have been trying to crack for decades, which means MuZero could be a game-changer in the field.
Most game-playing AIs make their moves based on the most optimal outcomes. They plan ahead multiple phases, making contingencies for an opponent’s moves, then counter with the best possible outcome. The limitation here is that an AI playing like this can only make simple binary decisions, and can’t learn any more complex games, at least not without an overabundance of computing power. Instead of trying to perceive every possible outcome at once based solely on the rules, MuZero makes decisions based on its own circumstances. It calculates moves based on what just happened, what’s currently happening, and what it would like to happen next, not unlike how a human being plays games.
While MuZero hasn’t yet evolved beyond games of Ms. Pac-Man, this could be an important step in the development of genuine thinking computers. Hopefully, if they’re as free-wheeling as MuZero, we won’t have to worry about them stealing nuclear launch codes from Matthew Broderick.