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

[Henk]

Yes, it's slower. The advantage is that it may learn things beyond humans' imaginations, and for me that's a cause worth pursuing.

So it will be a competition between better evaluation versus an extra ply.

[matthewlai]

Yes, it's slower. The advantage is that it may learn things beyond humans' imaginations, and for me that's a cause worth pursuing.

So it will be a competition between better evaluation versus an extra ply.

I will also be working on tree-shaping later, also using neural networks. Ideally that would make the search more selective, and be able to reach deeper.

[Henk]

Yes, it's slower. The advantage is that it may learn things beyond humans' imaginations, and for me that's a cause worth pursuing.

So it will be a competition between better evaluation versus an extra ply.

I will also be working on tree-shaping later, also using neural networks. Ideally that would make the search more selective, and be able to reach deeper.

Perhaps also or only use those neural networks for move ordering. But not to close to the leaves otherwise you get another computational bottleneck.

[matthewlai]

Yes, it's slower. The advantage is that it may learn things beyond humans' imaginations, and for me that's a cause worth pursuing.

So it will be a competition between better evaluation versus an extra ply.

I will also be working on tree-shaping later, also using neural networks. Ideally that would make the search more selective, and be able to reach deeper.

Perhaps also or only use those neural networks for move ordering. But not to close to the leaves otherwise you get another computational bottleneck.

That's definitely something to think about.

[mvk]

the scores won't be comparable to other engine's scores, since it's probabilistic.

You can map probabilistic scores to pawn scores with the inverse of the logistic function: 4*log10(p/(1-p)), if p is between 0 and 1.

And, coincidently, if your probabilistic score is the output of a neuron, and if the step function of that neuron is the logistic function, then you might simply report the sum of that neuron's inputs...

[matthewlai]

the scores won't be comparable to other engine's scores, since it's probabilistic.

You can map probabilistic scores to pawn scores with the inverse of the logistic function: 4*log10(p/(1-p)), if p is between 0 and 1.

And, coincidently, if your probabilistic score is the output of a neuron, and if the step function of that neuron is the logistic function, then you might simply report the sum of that neuron's inputs...

Indeed. The reason I don't really want to do that is not technical.

I want to use probabilistic scores internally in my program because I believe it is more "correct". I can just change the score in the printing function, but as I need to look at those scores a lot during testing, it could become confusing for myself.

That said, I will probably do it anyways for the next update.

The activation function of the output neuron is hyperbolic tangent.

[modolief]

Do you need compute resources for testing?

[matthewlai]

Do you need compute resources for testing?

I may in the future, when I want to do more testing with other engines. I don't think it's quite at that stage yet.

I have been using a computing cluster for testing and training, but the way it's set up means it may be difficult to actually run other people's engines (and also the cluster is obviously Linux-only).

[styx]

very interesting engine! it managed to win some fast games against me

somehow fun to play for a patzer like me

great achievement, good work!

[matthewlai]

very interesting engine! it managed to win some fast games against me

somehow fun to play for a patzer like me

great achievement, good work!

Thanks for trying it out!

Isn't it nice when we can win against an engine with no handicap .