In October of this year DeepMind (the artificial intelligence company of Alphabet) surprised us all with AlphaGo Zero. This new version of his artificial intelligence had learned to play Go without the help of humans and in a matter of hours it was the best in the world. He has not settled for Go, now he is also the best in the world in chess and in shogi (Japanese chess). But her final goal is not to be the champion in all the board games , she is only training.
A new paper published by the company describes how the second version of AlphaGo Zero, AlphaZero, has learned to play other board games . For Go, she was given only the rules of the game, and in a few hours she had already beaten AlphaGo by 100 wins to 0. She trained alone until she was the best. With AlphaZero and chess something similar has happened, knowing the rules and training for four hours has had enough to beat Stockfish (one of the most powerful chess engines in the world). As if this was not enough, with two hours of training in the shogi has already become unbeatable.
The key is in self-learning
Why AlphaZero learns so many board games? It does not do it for fun, DeepMind engineers have a greater purpose , to be able to apply these learning techniques to many other areas. They are still far from being able to achieve this, since it would mean being able to expand and open AlphaZero’s thinking to such an extent that they are capable of learning anything (with many more variables than the rules of a game), like a human.
But AlphaZero and these “small victories” in table games are a first step. This artificial intelligence (it is a variant of AlphaGo Zero more generic) has been developed in such a way that it can learn by itself and not through the experience with external agents. AlphaZero can be applied to a wider range of fields and topics without having to teach it, so probably in the coming weeks we will see more progress and maybe not just in board games.
The evolution of DeepMind with its artificial intelligence engines is constant. In March of this year we saw how he was already able to remember and use what he had learned in new tasks. Subsequently, DeepMind showed us how they want to apply it to detect early signs of chronic eye diseases. He has also played Starcraft II and in October we saw how he learned to play Go on his own. I said, he’s only training.