First AlphaGo; Now AlphaZero: What DeepMind’s New Game-Playing A.I. Is Capable Of
By playing against itself, DeepMind’s latest game-playing artificial intelligence (A.I.), AlphaZero, has mastered not one but three games: chess, shogi, and Go. In an earlier iteration of the software, called AlphaGo, DeepMind researchers combined human guidance and machine learning to defeat the world’s best Go players. Now the software has proved itself able to master more than one game: a first according to DeepMind.
In a paper in Science Thursday, DeepMind researchers report that AlphaZero beat the best-known chess and shogi software programs, called Stockfish and Fritz.
In AlphaGo, programmers coded the rules of the game and some initial strategies known to humans. The game-playing algorithm then searched through possible future moves and their consequences using something called a Monte Carlo tree search (MCTS). The program stored what it learned from its searches and from playing test games in what’s called a neural network, named that way because it replicates some features of how human neurons store learned information in layers. Together with its massive processing power, the software began defeating the best human Go players in 2016.
Now the researchers have generalized the approach to two other important games, chess, and shogi, a Japanese variant of chess. While computers have been beating humans at chess since IBM’s Deep Blue beat Garry Kasparov in 1997, they have done it using a combination of stored memories of known chess strategies and brute processing power. The DeepMind approach instead requires that the software play a huge number of games against itself — more than humans have played in history — and storing what it learns to guide future gameplay.
In other words, it is given the rules of the game, then creates its own strategies, unhinged from human biases and reinforced only by victory or defeat.
“Explainability is still an issue — it’s not going to put chess coaches out of business just yet,” writes Kasparov. That’s a common issue across self-learning artificial intelligence systems. Still, computer scientists and their software brains may soon grow bored with chess, shogi, and Go, writes IBM Research computer scientist Murray Campbell in Science: “AI researchers need to look to a new generation of games to provide the next set of challenges.”
That’s something Theodore from the film Her might be able to understand.