The training is the really computationally intensive bit with neural nets - fairly sure that actually running them once trained doesn't take anything too crazy.

I genuinely didn't think this approach would work quite so well in chess - the brute force search is awfully effective of course. It seems like chess is a bit more interesting than we'd thought

Pretty impressive, even if there seems to be a question of the fairness of the Stockfish comparison. I'm inclined to believe they are not making this up based on this being the DeepMind team, who did create AlphaGo in a similar, if slightly more domain specific, way. It would be fascinating whether it does very differently than current engines somewhere.

bragesjo wrote on 12/06/17 at 19:42:25:


The DeepMind program was essentially "playing as a human". In other words it had no resource to external assistance other than what it had learned for itself.

On the basis of the games so far published, it has replicated human opening theory and confirmed some assessments without suggesting anything dramatically new.

https://chess24.com/en/read/news/deepmind-s-alphazero-crushes-chess here's a more detailed easier readable article on the topic. It used neither opening books nor tablebases, but just played games against itself until it was the best, reinventing the past centuries of opening theory on the way  

Disclaimer: Very unclear basis on everything right now!

According to its developers, AlphaZero (as far as I understand it, the updated version of AlphaGo, the Go program which beat a human elite player for the first time in history) learned chess "in 4 hours", then played a 100 game match with Stockfish 8, and scored +28-0=72 (including 3 wins with black)!

The time control was supposedly 1minute/move, with AZ pulling farther away from Stockfish the longer the timecontrol happened to be.
It uses a drastically different algorithm; rather than going through 70 million positions per second (Stockfish), it checks just 80 thousand, but uses its "deep neural network" to focus much more efficiently on more promising variations (plays like a human, so to speak).

If all of this is true, and it ever makes its way into human hands (currently its not exactly available, and very unclear how well it will fare if it eg were to use less than 64 cores), that would be big news indeed

https://lichess.org/study/EOddRjJ8 You can check out 10 sample games of the 100 game match here (G7+G8 are two of the three Black wins)
https://arxiv.org/pdf/1712.01815.pdf if you want a more detailed explanation of .. everything

Right now it looks a vague bit suspicious to me, as we have no independent party confirming anything, and the games seem a bit strange -
https://lichess.org/0LUhNlLB here for example is one of Stockfish's White losses, which the quick Lichess Analysis (also running Stockfish) evaluates as having 5 Inaccuracies + 1 Mistake & 15 ACPL; basically unheard of in elite Engine chess. That it may misevaluate the AlphaZero program if that is indeed stronger than itself I can understand, but I see little reason as to why Stockfish would criticize *its own* moves to that extend (to compare: Analysis of G94 in Houdini vs Stockfish from TCEC 2016 https://lichess.org/Fp9UmSGv)