Author: Robert Hyatt
Date: 07:25:52 07/14/02
Go up one level in this thread
On July 13, 2002 at 02:37:06, Dann Corbit wrote:
>On July 13, 2002 at 02:23:00, eric guttenberg wrote:
>
>>Are you saying that a program that falls for a trojan horse sacrifice
>>(as most programs apparently do, at least at shorter time controls)
>>will eventually "figure it out" and stop accepting the sacrifice??
>>
>>I am aware that most programs have "book" learning but I do not know that
>>they can learn from positions. If so, would not Nemeth be discovering
>>this from the programs he is beating?
>>
>>Someone please correct me if I am wrong, but I thought chess programs
>>do not have positional learning.
>
>Many chess programs (including amateur programs) have positional learning.
>
>Examples include Crafty, Sjeng and KnightCap.
>
>Now, I should temper what I said. You might discover a principle that defeats a
>chess engine. Then you can possibly fool it again with transpositions or
>perhaps with applying the same principle but something changed a little bit.
>
>You might also find a bug in a program and exploit it over and over.
>
>Some programs do not learn (I think most {if not all} professional programs do
>learn). If they do not learn, they certainly ought to.
>
>It might also take quite a while for enough statistics to accumulate to know
>that a particular move is bad.
>
>Suppose (for instance) that a chess knows position Q wins 300, loses 200 and
>draws 100. You may have to beat it 100 times before the position looks even if
>the program only has book learning. So for a long time you can beat the
>program. But at some point enough statistics accumulate and the attack stops
>working.
THat is the point. Everyone should read Slate/Scherzer's article on "learning
using the hash table" in an old JICCA. This was what I based my position
learning on. In the article, Dave shows a position where the program really
wants to capture a hung piece. But doing so opens a diagonal and leads to a
smothered mate. There are a couple of ways to make the capture, and each
fails. And gets stored in the permanent hash (position learning file). After
a couple of runs, it figures out it can't make the capture and it therefore
does the "right thing" and doesn't end up getting mated..
It shows how position learning helps for _specific_ traps. It fails miserably
for "theme" traps however.
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.