On December 5th a research paper was published titled Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm. In the paper, it described a match between computer chess engine Stockfish and DeepMind's AlphaZero.
Stockfish is a giant amoung open-source chess engines. It is the 2016 Top Chess Engine Champion (TCEC) and can evaluate around 70 million positions per second.
AlphaZero on the other hand, can only evaluate 80,000 positions per second. Oh, and AlphaZero just needed 4 hours to play itself from not knowing the rules of chess to being able to say "ok Stockfish, let's play and see how I do."
Just viewing the opponents on paper, I'm not sure anyone would have guessed the results of a 100 game match. AlphaZero won 28 games and the rest was drawn.
Let this sink in: Stockfish got skunked and didn't win a single game. It was similar to Magnus Calsen accepting a challenge from Max Deutsch that he could learn chess in a month and beat him. Max learned a tough lesson that day: only a computer program can learn and compete at a grandmaster level in 30 days.
AlphaZero's team says it uses an algorithm through a deep neural network that efficiently focuses what it thinks are the most promising position variations.
I'm not going to pretend to know what the real technical sorcery is, but feel free to read the paper and this article.
Things to consider:
- not sure if paper was peer-reviewed and/or independently checked.
- is it a fair match if the hardware that the engines are running on are not equivalent? I'm not clear if either engine could be run on the other's platform.
Regardless, the results, are stunningly lopsided.
I thought chess engines were basically an area these days where improvements were marginal. Its fascinating and a little scary to know that there was more room for improvement.
Even if the match was a little unfair, AlphaZero's algorithm may accelerate other areas of AI and broaden it's application in a variety of business areas.
Indeed this is a great time to be alive.
These posts have greater value when you comment. Please share your thoughts with the Steemit community.
Traversing the Cryptosphere,
Kryptonaut
