Christopher wrote:
Robert Hyatt wrote in message ...
Christopher wrote:
Hi!

Does anyone know how I can test Crafty learning?

I mean. What chess problem can I enter .. wait.. and know that Crafty
learned something when it say's "learning" when I hit new or quit. I
see Crafty is indicating that it has learned something but I want to
test to see what was learned or what really happened.

I guess, I want to know how to verify it. I entered a game. Crafty
said it learned something then I play the same position that is said
it learned and it seems that the speed and results are the same.

What I am doing is inputting chess problems but when Crafty says it
learned something it seems to do the same thing.

Find a chess position Crafty can't solve from the root. Let it search
and make a move. You make a move (a good move that refutes the move
it played) and let it search to see that move was bad.

Start over and it should play a different move (maybe not the right
one yet, but not the first "wrong one."

That's how it works, by using results form later searches to get more
accuracy earlier.



Maybe I just don't understand what the learning function is or is
there a way I can test do demonstrate that it has learned something.
I guess I am looking for a different result so I can say .. yeah
before it did this... now it's doing something different(better or
worse isn't important at this stage. just want to know how to test
it).

Any thoughts or suggestions are most appreciated.
Kudos to Dr. Hyatt!

Thank you very much for the help. I will try that as you suggested.
If I can, how does Crafty know that the move was refuted? Is it
because of a loss in the game? Or is it because the predicted value
for its original move lost points?

The idea is this. Crafty makes a move (Move 21) in the game with
a search score of +.47... When it gets to move 22 after the opponent
makes a move, it searches and now sees a score of -.21... It stores
this -.21 in the "permanent position learning file". If you start
over and reach move 21, it may well be able to use that information
from move 22 to find a _better_ move at move 21. It _may_ but it
isn't guaranteed as the tree is very bushy. But that's the idea.



ex: Crafty makes a move(crafty thinks it's value is .5). I find a
move better. Then crafty on the next move sees that it's value went
to (.25) then is that when it remembers that move was poor? Or does
is wait for 5 moves later or does it learn when the game was won or
lost?

I ask because when I give it a problem and let it think for 1 hour it
makes a move. Then I quit the game right after and Crafty says
"Learning position" (something to that affect). I am curious to know
what it learned if I didn't respond.


That just means the score was probably well below zero, which is bad.




I guess I was thinking I could give it a bunch of games and let it
think and when I give it the same position it should do something
better or faster. Maybe that is where I am not understanding.

Otherwise would it make sense if I gave Crafty a bunch of games and
let another program think for more time(Fritz8 or Jr8) and then when
Crafty plays again the same line Crafty should remember and play
something different(I understand not necesarily better but different).
I am trying to understand the function so if I let it play Fritz8
that Crafty should get better (assuming Crafty is on a PC slower than
Fritz).

Also can I get Crafty to play Fritz without me having to enter the
moves?

Thanks Much Dr. Hyatt

That should be possible. I think the fritz GUI will play games between
Fritz and other engines, including crafty.





--
Robert Hyatt Computer and Information Sciences
University of Alabama at Birmingham
(205) 934-2213 115A Campbell Hall, UAB Station
(205) 934-5473 FAX Birmingham, AL 35294-1170