Hacker News new | past | comments | ask | show | jobs | submit login
DeepChess: End-To-End Deep Neural Network for Automatic Learning in Chess [pdf] (tau.ac.il)
61 points by lainon on April 15, 2017 | hide | past | favorite | 11 comments



This is an interesting idea, but the authors lose credibility by using Crafty as an opponent. Crafty isn't even among the top 30 current programs. Depending on which rating list you reference Crafty is 300 to 400 points worse than Stockfish.

>Crafty has successfully participated in numerous WCCCs, and is a direct descendant of Cray Blitz, the WCCC winner of 1983 and 1986.

That is some pretty weak sauce. "successfully participated"? Is that really worth mentioning? Also Crafty isn't exactly the direct descendant of Cray Blitz even though they were both written by the same person. And while Cray Blitz did win a World Championship some 30 years ago, it did so on the power of its hardware, not the strength of its software. It was running on a Cray supercomputer at a time when other programs were running on vastly inferior hardware. Rebel, who finished one point behind Cray Blitz, was running on a 6502.

Section 5.1, where they show diagrams on moves made by DeepChess is only meaningful if other programs miss the moves.

>This property has been associated with human grandmasters, and has always been considered an area in which computer chess programs were lacking.

But...

In the Aronian game Stockfish running in the browser (PNACL) finds the move Re5 in less than a second.

In the Alekhine game Stockfish finds the move d5 in about 2.5 seconds.

In the Seirawan game Stockfish finds the move c5 in about 1 second.

So while I find the idea interesting I'd like to see more intellectual honesty in this type of paper.


But..but.. Crafty still has an ELO of 3030, which is only about 300 pts below Stockfish. It would beat most human chess players (most grandmasters too), which is still quite an achievement, given that this neural net learned to play chess by itself (using no such human derived evaluation function). Cool right?


To put things into context, the best grandmasters in the world have a rating around 2800.


Beating Crafty is enough to prove that it can beat any human.. that is quite an achievement, or atleast a notable result. Of course, its easy to criticize without having played chess or doing any research. Going by Crafty's ELO, it would win against Magnus Carlsen 80% of the time


> Going by Crafty's ELO, it would win against Magnus Carlsen 80% of the time

Careful with that logic, I'm not sure that engine ELOs are in the same rating pool as human players (and so rating comparisons between the two are sort of meaningless).


carlsen (like all super grandmasters) would happily concede that he can't beat 3000+ engines most of the time. i don't know if chess commentators appreciate what this means about chess engines and the fundamental principles of chess (i know i don't / can't), but it's universally agreed that these "3000 ELO" chess engines are demonstrably stronger than every known human chess player.


Why would it ever lose to Carlsen?

When was the last game an ELO > 3000 AI lost to a human at all?


"Elo's central assumption was that the chess performance of each player in each game is a normally distributed random variable."

https://en.wikipedia.org/wiki/Elo_rating_system


And in the Tal vs Larsen game, it looks like the sacrifice may not be 100% sound.

http://www.mark-weeks.com/aboutcom/aa05e21.htm


I found this implementation using TensorFlow: https://github.com/mr-press/DeepChess


It is interesting that they removed they didn't allow positions that resulted from captures in the training dataset. I wonder why would that affect negatively the results. The explanation they give (opponent will likely capture back soon) does not sound enough to me. The tactics of the position should capture that. Is this because it is hard for the network to learn from positions with less material and tactical opportunities?




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: