*First release* Giraffe, a new engine based on deep learning

[stegemma]

Today I've done a little match between Satana and Giraffe. You can find the games on my website www.linformatica.com

The style of Giraffe is very interesting... and strange almost as those of satana

[matthewlai]

Today I've done a little match between Satana and Giraffe. You can find the games on my website www.linformatica.com

The style of Giraffe is very interesting... and strange almost as those of satana

Thanks! The games look pretty interesting . One big problem with Giraffe right now is that she always wants to get the king out early. That's probably positions in endgames are trained faster (because they are close to the final reward), so the eval is biased for endgames, where the king should get out.

It also really likes to push pawns, probably for the same reason.

[stegemma]

Today I've done a little match between Satana and Giraffe. You can find the games on my website www.linformatica.com

The style of Giraffe is very interesting... and strange almost as those of satana

Thanks! The games look pretty interesting . One big problem with Giraffe right now is that she always wants to get the king out early. That's probably positions in endgames are trained faster (because they are close to the final reward), so the eval is biased for endgames, where the king should get out.

It also really likes to push pawns, probably for the same reason.

Satana has similar problem: it moves the queen too soon and rarely it play castling. But don't worry, even young players starts this way... so your neural network is working like a child's brain!

[matthewlai]

Today I've done a little match between Satana and Giraffe. You can find the games on my website www.linformatica.com

The style of Giraffe is very interesting... and strange almost as those of satana

Thanks! The games look pretty interesting . One big problem with Giraffe right now is that she always wants to get the king out early. That's probably positions in endgames are trained faster (because they are close to the final reward), so the eval is biased for endgames, where the king should get out.

It also really likes to push pawns, probably for the same reason.

Satana has similar problem: it moves the queen too soon and rarely it play castling. But don't worry, even young players starts this way... so your neural network is working like a child's brain!

That is always good to know .

[yanquis1972]

is it based on learning at an end user level, or is that simply how its developing on your end?

ie will it learn as i play it or run it against other engines?

[matthewlai]

is it based on learning at an end user level, or is that simply how its developing on your end?

ie will it learn as i play it or run it against other engines?

It is learning on my end only. There's no learning through normal gameplay.

[yanquis1972]

thanks -- very cool to see someone taking a novel approach to engineering a chess program. the high level engines are great but i definitely think theres a 'market' for interesting and unique engines as well.

[matthewlai]

thanks -- very cool to see someone taking a novel approach to engineering a chess program. the high level engines are great but i definitely think theres a 'market' for interesting and unique engines as well.

Yeah it's a lot of fun .

With conventional approaches I'm always out of ideas to try, and most of the ideas I do try do no better than existing approaches, which is to be expected since existing approaches have been fine-tuned for decades already.

With this new approach almost everything is still unexplored. I have a huge list of stuff I want to try, and quite a few of them have resulted in good improvements. I can't implement them fast enough!

[clumma]

Very interesting project. I have two questions:

1. Why did you bootstrap from material-only eval? Doesn't the CCRL dump contain evals by the best engines? Why not train on those?

2. Are you familiar with the idea of model compression -- training a small network to mimic a larger one? E.g.
http://arxiv.org/abs/1312.6184
http://arxiv.org/abs/1503.02531
and do you think this could be used to speed up Giraffe's eval?

Cheers,

-Carl

[melajara]

One big problem with Giraffe right now is that she always wants to get the king out early. That's probably positions in endgames are trained faster (because they are close to the final reward), so the eval is biased for endgames, where the king should get out.

Could you not compensate for this training bias by providing more context parameters, actually modelling what we, as humans, understand as the 3 game phases in chess (opening, middlegame, endgame)?

Doing so, development moves (but not king move beside castling, lol) would be favored in opening and king centralization (a strong point in Giraffe's endgame play) reserved for endgame?