Author: Charles Roberson
Date: 10:35:46 03/06/06
Go up one level in this thread
On March 05, 2006 at 09:19:08, walt irvin wrote:
>On March 04, 2006 at 11:52:07, Charles Roberson wrote:
>
>>On March 04, 2006 at 10:52:52, walt irvin wrote:
>>
>>>On March 04, 2006 at 09:56:10, Charles Roberson wrote:
>>>
>>>>
>>
>> The math seems to be working against you on this.
>
>ok i agree the math is working against me on that idea,,,,,what about getting a
>huge database of already played games and use that to learn with ???? that
>should not take 20 years ???? im sure there are many many many computer games
>out there and for that matter gm games that u could use to help your program
>learn faster ,,,,
>
>walter irvin
That is more reasonable, but there are still issues. GM's and computers
make mistakes. Thus, moves in the database need verification but maybe
we drop the idea of perfection and go for "close enough".
Now, the big issue is how to do the learning. Several methods exist:
create a minimax tree, Bayesian trees, Neural Networks, .......
The issue with the tree methods is how to handle propogation of the leaf
value up the tree (minimax it or Bayesian). Both have their pros and cons.
But they are fundamentally limited by the "close enough" strategy.
The fundamental weakness in the "close enough" strategy is:
Given a GM match that was lost by (say black), the methods will degrade
moves leading to that position. However, what if black made a very subtle
mistake in the late game and a computer or another GM could find that
it is a forced draw or a forced win for black?
Even this is not simple. But it is doable in many different ways (some bad,
some not so good, some good, some great). That is where the research is.
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.