designing neural networks

[Rémi Coulom]

Hi Gerd,

For what it's worth, I have talked with Gian-Carlo Pascutto in Amsterdam, and he was telling everyone about his succesful machine-learning experiments. If I remember correctly, he used no hidden layer. Evaluation function was a weighted sum of features, and the learning algorithm trained the weights. Hey, but you were there too! He probably told you already.

If you wish to study neural networks, the very best book is that of Chris Bishop:
http://research.microsoft.com/~cmbishop/nnpr.htm

Regarding your question about network architecture, I would say it might not be such a good idea to really use a network. Neural-network algorithms are in fact, more generally, methods to tune the parameters of a function approximator. So, if you write an evaluation functions that computes a value from the position + parameters, you can use those machine-learning algorithms to tune the parameters directly, without really having a network that looks like a multi-layer perceptron. The only property that you need is that you should be able to compute the derivative of the evaluation with respect to parameters, and maybe second-order derivatives too.

Good luck with your experiments.

Rémi

[Gerd Isenberg]

Hi Gerd,

For what it's worth, I have talked with Gian-Carlo Pascutto in Amsterdam, and he was telling everyone about his succesful machine-learning experiments. If I remember correctly, he used no hidden layer. Evaluation function was a weighted sum of features, and the learning algorithm trained the weights. Hey, but you were there too! He probably told you already.

If you wish to study neural networks, the very best book is that of Chris Bishop:
http://research.microsoft.com/~cmbishop/nnpr.htm

Regarding your question about network architecture, I would say it might not be such a good idea to really use a network. Neural-network algorithms are in fact, more generally, methods to tune the parameters of a function approximator. So, if you write an evaluation functions that computes a value from the position + parameters, you can use those machine-learning algorithms to tune the parameters directly, without really having a network that looks like a multi-layer perceptron. The only property that you need is that you should be able to compute the derivative of the evaluation with respect to parameters, and maybe second-order derivatives too.

Good luck with your experiments.

Rémi

Hi Rémi,

yes, I know Gian-Carlo is quite convinced with his learning approach. One further reason I am interested in. I am not aware on implementation details though. I thought the hidden layer was somehow required for some learning algorithms. What do you think obout using an explicit sigmoid function e.g. to scale endgame-middlegame terms on sum of material?

Thanks for your wishes,
Same for you as well.

Gerd

[Rémi Coulom]

Hi Rémi,

yes, I know Gian-Carlo is quite convinced with his learning approach. One further reason I am interested in. I am not aware on implementation details though. I thought the hidden layer was somehow required for some learning algorithms. What do you think obout using an explicit sigmoid function e.g. to scale endgame-middlegame terms on sum of material?

For endgame you would write something like

Code: Select all

V += EndGameTerm * sigmoid((Phase - Threshold) * Steepness)

, where Threshold and Steepness are two parameters of your evaluation function. The learning algorithm would then tune Threshold and Steepness. There is no real need to create a full-blown neural network for this.

This should have very little cost in terms of computation time. If you wish to avoid the cost of floating point operations, once the parameters have been tuned, maybe you can create a lookup table indexed by Phase. But you probably know the optimization tricks much better than I do

Rémi