Hi.

Apparently a 1000 game match has been played between Alphazero and Stockfish 8. Here are a few tidbits.

This seems like a case of Alphazero overplaying his hand. Possibly Alphazero miscalculated before playing his 79th (?)


Alphazero wins with the Dutch (24...Ng5!!)


And doesn't mind playing (and winning with) the Berlin:


Have a nice day.

gillbod wrote on 04/05/18 at 19:50:40:


Sorry I missrembered. We actually dont know exactly what hardware Stockfish used, the paper only said the number of threads (64) and 1 GB hashtable. Still Alpha Zero had better hardware but not 2000 times better as I stated.

I am no expert but I read at some forum that a single TPU has the same theoretical power as 500 Haswell cores. I thought I used the same numbers as during training but in the match it used 4 TPUs. And 4 TPUs means the power of 2000 Haswell cores so that where I missrembered the numbers from.

Don't forget that just doing this a bit later would have generated tremendously different results -
The devbuild of SF9 is currently up +11-1=27 vs Houdini 6, incl. a few wins with Black and most notably a "double-win" (winning the same position with both colours). 
https://chess24.com/en/watch/live-tournaments/tcec-2018-superfinal/1/1/30

So yeah, overall this was mostly a proof-of-concept paired with an ad campaign. Of course an important proof, and I look forward to what Leela will produce in the future, but it's been rather thinly veiled that Google doesn't care about chess whatsoever and just wanted to score a nice headline in time for NIPS, and most likely consciously fiddled with the SF settings until they were to their liking.
Still doesn't matter too much to me - if A0 against "real" SF had gone even or potentially even lost, it would still just have been a question of more time (and/or better optimization for chess) until A0 has improved enough to 'really' beat Stockfish, so in the end it's more a bending of truth than a lie.

I don't think any of those things make the match 'fixed'. They mean that Alpha Zero wasn't playing against the strongest possible Stockfish that humanity could have been assembled.

Personally I very much doubt that a Stockfish with all those changes could have beaten the version that did play by a +28 score over 100 games, but who knows? Even if it could it really doesn't matter and is besides the main point. Alpha Zero taught itself to play chess with only the basic rules as an input and is a major breakthrough.

tdv wrote on 03/28/18 at 18:40:53:


Why? Stockfish can't run decently on GPUs, why insist that Alpha Zero should run on CPUs?

Also, it's perfectly possible (even relatively easy) to set up a GPU instance in AWS infrastructure and have it talk to your local chess GUI (i.e. the heavy lifting will be done in the cloud, but you can still use your chess program of choice on your home computer). 

So the hardware that Alpha Zero ran on is perfectly accessible to most dedicated people at home without having to buy a new computer.

tdv wrote on 03/31/18 at 02:07:08:

Leela Zero (the Go program) is now pro strength.

https://www.reddit.com/r/chess/comments/882jr8/leela_chess_zero_the_fork_of_leel... There is now an effort to 'recreate' AlphaZero in Open Source! It's still quite a bit away (the post mentions ~400 days at current pace), but still quite interesting & not impossible that it will pick up some speed with more people joining in.
You can also play against the engine at its current level, and from what it sounds like (I haven't tried yet), it plays a lot more like actual humans than a handicapped engine normally would, which is exciting stuff.

fjd wrote on 01/19/18 at 19:30:30:


Also covered by John Emms here on ChessPub, of course: http://www.chesspublishing.com/content/10/index.htm