DeepChess? Another deep-learning based chess program

[matthewlai]

Has anyone seen this paper?

http://www.cs.tau.ac.il/~wolf/papers/deepchess.pdf

I stumbled upon it today. It's by Omid (Falcon author) et al, about using a comparison-based alpha-beta with a deep network. This is something I really wanted to explore but didn't have a chance to.

They claim it's 60 Elo stronger than Crafty, but like Falcon, I don't think the code (or the binary) has been released.

[mar]

They claim it's 60 Elo stronger than Crafty

Unfortunately they claim so based on 100 games...

Still pretty impressive. Would be nice to for people to test the binaries.

[bob]

Has anyone seen this paper?

http://www.cs.tau.ac.il/~wolf/papers/deepchess.pdf

I stumbled upon it today. It's by Omid (Falcon author) et al, about using a comparison-based alpha-beta with a deep network. This is something I really wanted to explore but didn't have a chance to.

They claim it's 60 Elo stronger than Crafty, but like Falcon, I don't think the code (or the binary) has been released.

They also don't explain the circumstances of the 100 game match vs Crafty. Book? No Book? Etc.

It really is pretty sloppy to give 100 game results today, when most play tens of thousands of games to make tuning decisions...

[Rochester]

They claim it's 60 Elo stronger than Crafty

Unfortunately they claim so based on 100 games...

Still pretty impressive. Would be nice to for people to test the binaries.

You can say the progression not originate from the inside people. They are thinking the same ways always. Only small change of the same thing. Forever small change. But the revolution come from the outside! The new mind can beat the old programme even when work 50 years on the same thing. No new idea in the first 10 years, then never the new idea in the next years also. Mr Hinton is the hero for the computer science! He already invent this long time ago. Only now the computer have the good speed. First the go, now the chess.

[CheckersGuy]

They claim it's 60 Elo stronger than Crafty

Unfortunately they claim so based on 100 games...

Still pretty impressive. Would be nice to for people to test the binaries.

You can say the progression not originate from the inside people. They are thinking the same ways always. Only small change of the same thing. Forever small change. But the revolution come from the outside! The new mind can beat the old programme even when work 50 years on the same thing. No new idea in the first 10 years, then never the new idea in the next years also. Mr Hinton is the hero for the computer science! He already invent this long time ago. Only now the computer have the good speed. First the go, now the chess.

However, go and chess are completely different games. Go is much more strategical than chess. Ther reason researchers turned to neural Networks for solving the game of go was because in go it's hard to create a static Evaluation function that tells you how the game may end. I think it will be quite a while until we see a neural Network based Evaluation function in one of the top engines because they are way too slow and impractable today.

[Dann Corbit]

http://senseis.xmp.net/?ComputerGo

[kinderchocolate]

I've read the paper, but I don't quite understand:

Q1: How does the binary output representation create the evaluation function? In Giraffe, the weights for the evaluation are the parameters, but I don't see anything like that mentioned in the paper.

Q2: What's the new alpha-beta algorithm? It looks very much like the classical method.

Q3: Why training a one-hot encoding is more accurate than a binary output?

Q4: How does the model deal with two equivalent moves, for example, two different ways to get a checkmate in 5?

Q5: Has the paper been peer-reviewed?

[matthewlai]

I've read the paper, but I don't quite understand:

Q1: How does the binary output representation create the evaluation function? In Giraffe, the weights for the evaluation are the parameters, but I don't see anything like that mentioned in the paper.

They don't have an evaluation function in the traditional sense. Everything is based on comparing pairs of positions.

Q2: What's the new alpha-beta algorithm? It looks very much like the classical method.

They use position comparisons. So alpha and beta are both positions instead of scores. Positions don't have scores.

Q3: Why training a one-hot encoding is more accurate than a binary output?

Do they actually use one-hot? In any case I expect them to be the same. If you only have two outputs, softmax one-hot becomes logistic.

Q4: How does the model deal with two equivalent moves, for example, two different ways to get a checkmate in 5?

My guess is the return value will be arbitrary. It wouldn't really matter.

Q5: Has the paper been peer-reviewed?

I believe so, but the code wasn't released, so the reviewer(s) probably did not have access to the actual program.

[whereagles]

Cool. I'm sending this to a friend who's into neural networks and machine learning stuff.

[brianr]

An implementation:

https://github.com/mr-press/DeepChess

As an aside, please consider Keras.
I greatly appreciate all those who share their knowledge (and code).
However, as a machine learning newbie with aspirations of someday being a novice (a bit like Carl Spackler's six year plan[http://www.imdb.com/title/tt0080487/), I hope to see greater use of Keras. Keras eliminates much of the pesky TensorFlow v Theano, Python 2 v 3, CPU v GPU, and Linux v Windows background noise.
Even I can get Keras to run.
Thanks again.