English Chess Forum

MartinCarpenter wrote: ↑

Thu Oct 19, 2017 8:12 pm

The problem with doing chess this way is more that the brute force approach is so very, very good at the game. Not much incentive to try and get fancy.

Seems like there was incentive enough. And Martin, no disrespect, but your assessment that the "brute force approach is so very, very good at the game" seems to be just wrong (unless the comments about a disparity in hardware are correct). Maybe in the future the fact that Stockfish was able to achieve 72 draws in 100 games (despite the hardware being different) will be seen as a mark of how good the current computer programming approach was. There will need to be new terminology to refer to the different computer-based approaches. Before AlphaGo and After AlphaGo.

Christopher Kreuzer wrote: ↑

Wed Dec 06, 2017 12:08 pm

(despite the hardware being different)

It would be interesting to know more about this. I have to admit that I am hunting for some reason why this shocking development is not quite as it appears.

NickFaulks wrote: ↑

Wed Dec 06, 2017 12:25 pm

Christopher Kreuzer wrote: ↑

Wed Dec 06, 2017 12:08 pm

(despite the hardware being different)

It would be interesting to know more about this. I have to admit that I am hunting for some reason why this shocking development is not quite as it appears.

Am only basing this on the Twitter comments such as this one:

https://twitter.com/esotericpig/status/ ... 4827381760

Bradley Whited on Twitter wrote:While impressive, seems like a PR stunt and a disrespectful slap in the face to the Stockfish developers. How can you compare TPUs to CPUs/GPUs? Hardware was skewed. And w/o an opening book, seems like a deep-machine-learning program would always start ahead.

TPU = Tensor processing unit.

NickFaulks wrote: ↑

Wed Dec 06, 2017 12:25 pm

Christopher Kreuzer wrote: ↑

Wed Dec 06, 2017 12:08 pm

(despite the hardware being different)

It would be interesting to know more about this. I have to admit that I am hunting for some reason why this shocking development is not quite as it appears.

There's a paper which says AlphaZero was evaluating 80K positions a second while Stockfish was looking at 70 million.

Christopher Kreuzer wrote: ↑

Wed Dec 06, 2017 12:38 pm

Am only basing this on the Twitter comments such as this one:

https://twitter.com/esotericpig/status/ ... 4827381760

Bradley Whited on Twitter wrote:While impressive, seems like a PR stunt and a disrespectful slap in the face to the Stockfish developers. How can you compare TPUs to CPUs/GPUs? Hardware was skewed. And w/o an opening book, seems like a deep-machine-learning program would always start ahead.

But did Stockfish have access to an opening book - and tablebases?

Angus French wrote: ↑

Wed Dec 06, 2017 12:45 pm

NickFaulks wrote: ↑

Wed Dec 06, 2017 12:25 pm

Christopher Kreuzer wrote: ↑

Wed Dec 06, 2017 12:08 pm

(despite the hardware being different)

It would be interesting to know more about this. I have to admit that I am hunting for some reason why this shocking development is not quite as it appears.

There's a paper which says AlphaZero was evaluating 80K positions a second while Stockfish was looking at 70 million.

The games AlphaZero (as it should more correctly be called) 'lost' in its self-training phase against Stockfish when constrained to follow popular openings were mostly as Black in the Sicilian Defence.

Sounds about right

Christopher Kreuzer wrote: ↑

Wed Dec 06, 2017 12:08 pm

MartinCarpenter wrote: ↑

Thu Oct 19, 2017 8:12 pm

The problem with doing chess this way is more that the brute force approach is so very, very good at the game. Not much incentive to try and get fancy.

Seems like there was incentive enough. And Martin, no disrespect, but your assessment that the "brute force approach is so very, very good at the game" seems to be just wrong (unless the comments about a disparity in hardware are correct). Maybe in the future the fact that Stockfish was able to achieve 72 draws in 100 games (despite the hardware being different) will be seen as a mark of how good the current computer programming approach was. There will need to be new terminology to refer to the different computer-based approaches. Before AlphaGo and After AlphaGo.

Well the brute force engines were good enough to utterly crush us! Rather astonished by this.

Its fascinating, but in a strange way this is a beyond the grave justification for the intuition of all the people who tried to produce 'intelligent' chess engines. Botvinnik as I remember?

Frankly I'm glad. Brute force chess engines were always a horribly ugly approach with multiple hacks like opening books required. This is so much cleaner and vastly more human like in how it works.

Seems like it really was a good problem after all

Can someone move this into one of the actual chess sections of the forum, please? Or create an Other Games & Computer Chess Section. Seems a bit rubbish dumping Go, amongst the other random dross.

Amusing that parallel architectures have gone past merely vector processing and now deploy Tensor Processing Units.

Paul McKeown wrote: ↑

Wed Dec 06, 2017 11:15 pm

Can someone move this into one of the actual chess sections of the forum, please? Or create an Other Games & Computer Chess Section. Seems a bit rubbish dumping Go, amongst the other random dross.

I agree, this doesn't belong in Not Chess. Computer Go is in any case the wrong thread, as this time the program was being used to learn chess.

Perhaps an Admin could create a new topic called (say) Machine Learning and move everything over to that?

I trust people saw that one of the authors was Dharshan Kumaran?

I certainly think we should move the last chunk of this thread. I'm not sure I'd want to lose the Alpha Zero stuff among everything else on the thread: it has huge implications and perhaps should stand alone.

The reason for the tensors is (I believe) that training the neural nets involves lots and lots and lots of matrix multiplication....

There's a really enormous technological fight going on between pure custom stuff like Google's TPU's and NVidia sticking specialist stuff into the most recent compute based versions of their GPU's. They're either after money or maybe thinking that our eventual robot overlord(s) will favour those who provided their brains

As I understand it, the first iteration of AlphaZero should have been making completely random moves like this

http://chessboardjs.com/examples#5002

If that's correct how on earth did it get started? A random v random chess game will end in a draw virtually 100% of the time, and if the game ends in a draw then it's useless for learning anything.

Of course I don't really understand this, but surely it would take billions of games before it would even start to develop any simple strategy such as 'it's a generally a good idea to take the opponent's pieces'. The paper says there were 700,000 steps of self-play but it doesn't say how many self-play games per step, unless I'm missing something. Really want to see what the early iterations look like!

Today I read in a newspaper about a new chess program called AlphaZero

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

What sort of computer does it run on?

There is a current discussion about AlphaZero in the thread Computer Go