A Chess forum. ChessBanter

If this is your first visit, be sure to check out the FAQ by clicking the link above. You may have to register before you can post: click the register link above to proceed. To start viewing messages, select the forum that you want to visit from the selection below.

Go Back   Home » ChessBanter forum » Chess Newsgroups » rec.games.chess.computer (Computer Chess)
Site Map Home Register Authors List Search Today's Posts Mark Forums Read Web Partners

Crazy idea for neural net eval function



 
 
Thread Tools Display Modes
  #1  
Old August 6th 03, 05:46 PM
Noah Roberts
external usenet poster
 
Posts: n/a
Default Crazy idea for neural net eval function


So the goal is to train the number 1 neural net to become an expert at
making the first move, the number 2 to become the best at the first
response, ...etc


I don't know a lot, but in my opinion this won't work; well it might,
but not the way you want. After a few plies the games become so
different that a ply based network would have to be as general as one
that simply plays from the start. So many different game states can be
reached at say ply 10 - is there a queen on the board, is it a closed or
open game, etc...


I was thinking about having 3-5 neural nets, opening, mid, and end, but I
thought, why stop there? The cutoff from opening, to midgame, to endgame is
pretty arbitrary.


What I would do is come up with a few specialized nets for specific
board properties. When some property changes you change the network.
Opening/mid/end is a decent place to start, but there may be more
interesting properties that you can use....castling, material, pawn
structure...


I'm basically hoping that the seperate neural nets will develop a one ply
lookahead ability that will help order the tree, like if you could ask a
three year old who understood the rules to point out the best move and
start there with the search.


Like I said, I don't think it could become an expert after a few plies.


A further improvement might be to somehow classify board positions as
belonging to a particular neural net using heuristics of some type.


And here is the one I would use. Ply just doesn't say anything about
the game.


Obviously this needs a lot of refinement, but I think it could have
promise.



  #2  
Old August 6th 03, 06:34 PM
Anders Thulin
external usenet poster
 
Posts: n/a
Default Crazy idea for neural net eval function



clameater wrote:

Obviously this needs a lot of refinement, but I think it could have
promise.


A couple of questions:

1. What is input to the NN, and what is output? (As precisely as
possibly)/

2. How do you propose to train the networks?

--
Anders Thulin http://www.algonet.se/~ath

 




Thread Tools
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

vB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Forum Jump


All times are GMT +1. The time now is 09:47 AM.


Powered by vBulletin® Version 3.6.4
Copyright ©2000 - 2017, Jelsoft Enterprises Ltd.Content Relevant URLs by vBSEO 2.4.0
Copyright 2004-2017 ChessBanter.
The comments are property of their posters.