How is AlphaZero different from Stockfish or Rybka?


I don't know much about AI or chess engines, but what is the fundamental difference between AlphaZero and Stockfish or Rybka?


Posted 2018-11-22T01:19:33.413

Reputation: 375



Stockfish is free software (source while AlphaGo can only be used by employees at Deepmind. This is important, because it means all answers to this question can only rely on the AlphaGo paper. Some decisions during the matches (e.g. giving limited time per move and not per game; giving different computational power) lead to unclarity how good alpha go really is.

Stockfish uses Alpha-beta search with a hardcoded metric to evaluate how good a position is.

In contrast, AlphaGo uses Monte Carlo tree search and convolutional neural networks. They are trained in an reinforcement learning setting.

AlphaGo is promoted as being almost directly applicable to different settings (Go, Chess), while stockfish is only applied to Chess. The techniques of stockfish can also be applied to Go, but the problem is the board evaluation. For Chess it's easy compared to Go.


Martin Thoma

Posted 2018-11-22T01:19:33.413

Reputation: 1 023

1Your references to AlphaGo should be to AlphaZero, although AlphaZero is derived from work on AlphaGo, and AlphaGo Zero. – Neil Slater – 2018-11-22T08:02:22.980