Author: Vincent Diepeveen
Date: 14:38:01 07/06/03
Go up one level in this thread
On July 06, 2003 at 16:21:05, Uri Blass wrote: >On July 06, 2003 at 15:42:25, Vincent Diepeveen wrote: > >>On July 06, 2003 at 08:00:48, Uri Blass wrote: >> >>>On July 06, 2003 at 03:04:07, Christophe Theron wrote: >>> >>>>On July 06, 2003 at 01:15:41, Uri Blass wrote: >>>> >>>>>On July 06, 2003 at 00:25:49, Uri Blass wrote: >>>>><snipped> >>>>>>>Maybe using it for the evaluation is not the most efficient use of a neural >>>>>>>network in a chess program. It seems that the way human players manage to search >>>>>>>the tree is vastly underestimated. >>>>>>> >>>>>>> >>>>>>> >>>>>>> Christophe >>>>>> >>>>>>I agree with you that search is underestimated in chess but I also believe >>>>>>that search and evaluation are connected because a lot of search decisions are >>>>>>based on evaluation of positions that are not leaf positions so you cannot >>>>>>seperate them and say search improvement gives x elo and evaluation improvement >>>>>>gives y elo. >>>>>> >>>>>>Uri >>>>> >>>>>I know that you did not try to seperate between them but my point is that if you >>>>>want to do the same as humans in the search then changing the search is not >>>>>enough. >>>>> >>>>>Humans may search position for some seconds and decide that this position is not >>>>>good and later search the same position but decide that it is good for them not >>>>>because they search deeper but because they learned to change their evaluation >>>>>based on searching other lines that leaded to a similiar position. >>>>> >>>>>Uri >>>> >>>> >>>> >>>>Well my point is just that when people talk about an application of ANN in chess >>>>they always talk about implementing the evaluation with an ANN, or tuning the >>>>evaluation with them. >>>> >>>>I think it tends to show that the application of ANN to chess has never been >>>>done by a "real" chess programmer. Because evaluation is only a part of a chess >>>>program. And maybe not the one that can be improved dramatically, or that needs >>>>them in order to be improved. Personally I would not use ANNs in the evaluation >>>>first, because I think they would be much more efficient somewhere else. >>>> >>>>On the other hand, you are right. If one could design an ANN to perform the >>>>evaluation, it would be wise to use the same ANN (or an extension of it) to >>>>guide the search. >>>> >>>> >>>> >>>> Christophe >>> >>>I believe that the biggest advantage that can be achieved in evaluation is not >>>in changing the initial static evaluation but in learning to change the >>>evaluation during the game based on the results of the search. >>> >>>I also do not believe that what humans know is the target and the target should >>>be better than what humans know. >>> >>>programs found better evaluation than humans in backgammon and program may find >>>better search rules than humans in chess not because programs are smarter but >>>because programs may do trillions of calculation to learn and humans cannot do >>>it. >>> >>>Uri >> >>This is the same utter nonsense crap that i keep seeing AI people write. Yet on >>average they even have less experience than you and keep believing in something >>they can never proof to be made. If they would have even *toyed* with ANNs a bit >>they will understand more about the impossibilities about it. > >I only say that I believe that it can be done. >It does not mean that I know how to do it. > >> >>Show me a backgammon program with an ANN that beats a 5 turns fullwidth >>searching backgammon program :) >> >>Of course show it at a machine that you and i have at home. > >Very easy >the 5 turns fullwidth searching backgammon program is going to lose on time >every game. > > > >> >>The average ANN expert is assuming he has to his availability something doing >>10^1000 calculations. > >I am not ANN expert and I did not suggest ideas how to do it. > >> >>That is the major problem when talking to these guys. >> >>Of course you can optimize an ANN for chess in 10^1000 calculations. >> >>But you will then be beaten by a database of just 10^43. >> >>I am however sure that 99% of all ANN interested will not understand what i >>write here above, simply because they do not know the running time of the learn >>methods applied. If they would read themselves into that, then less crap would >>leave their mouth. > >I did not say that the learning methods that are used in backgammon can work in >chess and it is possible that people need to invent different learning methods. >Uri If there was money to earn by programming a backgammon engine, i am sure some guys who are good in forward pruning algorithms like Johan de Koning would win every event there. It's like making a tictactoe program and then claiming that an ANN is going to work. As we have a saying here: "In the land of the blind, one eyed is King". That's why i focus upon chess. In contradiction to you, i know how to do it with ANNs (just like many others do), i just don't have 10^1000 system time to actually let the learning algorithm finish ;) Any approximation in the meantime will be playing very lousy chess... Hell, with 10^1000 runs, even TD learning might be correctly finding the right parameter optimization :) TD learning is randomly flipping a few parameters each time. It's pretty close to GA's in that respect.
This page took 0.02 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.