Author: Ingo Lindam
Date: 01:31:49 07/03/03
Go up one level in this thread
On July 02, 2003 at 20:38:09, Vincent Diepeveen wrote: >On July 02, 2003 at 12:44:56, Ingo Lindam wrote: > >> >>> b) they do not work very well for situations they are not trained for >>> and in chess you always explore new positions which are not trained yet, >>> which is an easy thing to understand once you understand that chess has >>> 10^44 positions and you could train perhaps for 10^2 positions at >>> most very well so missing around 10^40 somewhere. >> >>I am not voting for using NNs for chess, but is b) a fair agrument? >>Isn't it sufficient to see all significant features of positions often enough. >>Also a GM doesn't see 10^44 positions before he starts to play as a GM. >>Ofcourse the GM does obtain and evaluate some very concrete lines. And I don't >>suggest neither a human being nor a computer to play chess without calculating >>concrete lines. (As well as I would not suggest to play chess without having >>plans and aims and some kind of chess knowledge) > >Actually this problem is caused by 2 reasons > a) they get imperfect trained > b) all of the training algorithms are assuming the above b argument. > that's why NNs work for voice recognition a bit, as you usually have > the same voice, but they do not work for evaluatoin of chess positions > in playing programs because it is each time a new situation. > > It has no understanding like you and i do, all it has is a few > stupid parameters that it can tune. that's it. It doesn't even know > what a parameter influences in fact. It is all experiments that are > independant from each other. That's a real weakness trivially. Thank you Vincent, as already mentioned I am not a great fan of the idea to use NNs for chess. I would rather support the idea of an aproach using well defined statistical modells (that works pretty well for speaker independent large vocabulary speech recognition). Although some people might claim that NNs and a statistical approach could theoretically do the same. Ofcourse that approach has to be combined with a tree search and it might be a problem to find the right link between them and not to end in a pretty slow 3ply application. I could imagine a parallel system where the second processor (or some of the processors) are used to observe (parts) of the search space/tree to find patterns you have a usable kind of more specific knowledge about (or to define aims and goals or plans for parts of the search tree to cut/expand it). But also if this kind of knowledge would be not be usable by any engine/search algorithm in a sufficent fast way (I doubt about it) it would be a quiet helpful standalone tool for human beings. Internette Gruesse, Ingo
This page took 0.01 seconds to execute
Last modified: Thu, 07 Jul 11 08:48:38 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.