Author: Peter Fendrich
Date: 07:33:09 11/10/97
Go up one level in this thread
I don't think that all of the evaluation is suitable for NN's. Instead it might be better to "help" the NN's by specializing them for different tasks. Let's say King safety is one case. Decide what input parameters are useful for this case and train the net only with this in mind. The same for other evaluation components. With this aproach it also possible to use incremental updates of evaluation parameters. The big disadvantage is that spliting up the evaluation in different components like this will miss the interaction between the elements in different components. I think that there are lot of interactions between different elements on all levels and it will be extremely difficult to find the right components with the right input parameters without introducing this problem. So in the end I have my doubts about this but I'm still very curious of the results! Peter On November 10, 1997 at 09:48:36, Steve Maughan wrote: >I've some experience of neural networks and I don't think it would be >possible to generate a practical evaluation function with them. >Although neural networks can cope with non-linearity (eg a knight may or >may not be worth 3 points if positioned on e5), the huge range of >non-linearity involved in chess positions would defeat the non-linear >training algorithms. > >Another reason for shying away from neural Networks would be the time >required to evaluate a position. Due to the complexity of neural >networks they usually use floating point maths (but I guess that could >be substituted for lookup table) and require the calculation of >intermediate values (hidden layers). This would greatly slow down the >search compared to a linear evaluation function and would therefore need >to give a much more accurate positional evaluation to compensate. > >Also a complete evaluation will need to be carried out at the leaf nodes >since a non-linear element of evaluation may be altered by the last move >in the search chain (eg the last move Nf3 alters the value of the Bishop >on a7). This means that a program cannot perform incremental >evaluation, slowing it down even further. > >I don't want to sound too negative but I really think it would be tough >to develop a chess program based on Neural networks. However Go or >Othello are far more likely to be suitable! > >Regards > >Steve Maughan
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.