New Giraffe (Sept 8)

[PHALANX OF PAWNS]

Hello Carl,

The version I played was Giraffe version 20150908, so you are correct, must be an older version. Because it is a Winboard engine, I uploaded it into my old Chessmaster 10th edition platform. Of course, I gave it no openings book. Thank you for loading the game into an interface, but unfortunately I can't get it to work correctly!

[matthewlai]

How does Giraffe gets its chess knowledge ? Where does it come from ?

It's all described in the paper: http://arxiv.org/abs/1509.01549

[PHALANX OF PAWNS]

Hi Matthew,

You're absolutely correct, I am using the current version, but did not include the NN files. I will try this and report back!

[Henk]

What I understand is that you still have to tell Giraffe that it should use for instance Attack and Defend Maps. If so how do you know that that is right? .

[matthewlai]

What I understand is that you still have to tell Giraffe that it should use for instance Attack and Defend Maps. If so how do you know that that is right? .

Yes, but that's not required. It just makes the system perform a bit better.

I am actually thinking about removing that since I'm not sure if it's worth the speed lost.

[Henk]

What I understand is that you still have to tell Giraffe that it should use for instance Attack and Defend Maps. If so how do you know that that is right? .

Yes, but that's not required. It just makes the system perform a bit better.

I am actually thinking about removing that since I'm not sure if it's worth the speed lost.

Sliding piece mobility is the second (and last) redundant extra term you use. But you can not remove it I guess otherwise performance drops. That means or learning method is not quite optimal or input representation is not optimal or network representation is not optimal for ideally it should find all terms itself given position representation as input.

[But might be such optimal solution does not exist or too expensive to search for]

[Henk]

Perhaps it might even be that if you use material, side to move, mobility and one cell for network representation you get better performance.

[matthewlai]

Perhaps it might even be that if you use material, side to move, mobility and one cell for network representation you get better performance.

Better performance and much worse chess, yes. I don't think I want to do that.

[Henk]

Perhaps it might even be that if you use material, side to move, mobility and one cell for network representation you get better performance.

Better performance and much worse chess, yes. I don't think I want to do that.

Might even be better chess. No proof. No test. So I don't know. The less cells the easier to tune.

[matthewlai]

Perhaps it might even be that if you use material, side to move, mobility and one cell for network representation you get better performance.

Better performance and much worse chess, yes. I don't think I want to do that.

Might even be better chess. No proof. No test. So I don't know. The less cells the easier to tune.

I have tested extensively. The current network size is optimal.