DeepMind’s self-learning AI AlphaZero now easily plays three complex games

A system from Alphabet’s subsidiary DeepMind has mastered three of the most complex board games in the world. That’s what Silicon Angle reports. The artificial intelligence, AlphaZero, is the third major evaluation of AlphaGo, which three years ago defeated the world champion in Go.

It took months to train AlphaGo and it took a lot of human input to learn the difficult game. The second iteration, AlphaZero, was able to improve the level of its predecessor after 30 hours through a new learning method. The latest system also used this method to learn chess and Japanese shogi.

AlphaZero doesn’t look much like traditional computers for gaming like Deep Blue. Deep Blue won in 1986 from former chess champion Garry Kasparov. Traditional engines rely on information from human players, but AlphaZero has learned to play all by itself. The artificial intelligence was only given the rules of the three games it had to learn.

More powerful

“To learn every game, an untrained neural network plays millions of games against itself through a trail-and-error process called reinforcement learning. At first it plays completely randomly, but in the end the system learns from won and lost games, and adjusts the parameters of the neural network, giving it a better chance of making good choices in the future,” says DeepMind.

After only a few hours of training, the system was able to beat the former record holder in chess, Stockfish. In half of that time it managed to defeat the Elmo-engine, which was the record holder of shogi. The system has thus proven itself more powerful than other artificial intelligences.

According to DeepMind, the fact that AlphaZero was able to learn three different board games shows that the system may be able to master any information game. That category includes chess, Go and other games that hold the attention of researchers in the field of artificial intelligence.