Supercomputers have been beating the best human chess players for over 20 years, harking back to IBM’s Deep Blue in 1997. But, what has developed over the years with AI is the speed and efficiency of learning.
DeepMind, part of Google’s Alphabet, created the AI and trained it to learn the rules of chess in record time. It managed to beat the world champion chess program, Stockfish 8, in just 100 games. AlphaZero won or drew all of the 100 games, after playing itself over and over. It seems that this ‘self-play’ method of learning differs from the early days of supercomputers.
AlphaZero was only given the rules of the game of chess. Experts assumed that it took a lot more practical and ‘human’ knowledge to win, something that would take a lot more than 4 hours to learn. AlphaZero seemed to have a more human-like approach, searching for moves and managing to process around 80,000 moves per second, compared to Stockfish8’s 70.
(Super) powered by Google
However some experts are questioning the AI’s winning streak. Whilst much of the success was down to playing against itself, it had no other knowledge outside of the rules of the game. Was Google giving a little helping hand here?
DeepMind has access to a lot of computing power (thanks Google) – AlphaZero was trained on 64 TPU2s, compared to a much smaller 64 x 86 CPU threads for competitor Stockfish8. AlphaZero had a clear advantage here, training on super high tech hardware as opposed to normal PCs. Of course it would be able to learn faster thanks to all that power behind it.
So whilst it is awesome that AlphaZero learnt the game and became an expert in such a rapid amount of time, the experiment conditions need to be fair and easily judged first…