Author: José Carlos
Date: 12:26:59 07/29/04
Go up one level in this thread
On July 29, 2004 at 09:47:32, Uri Blass wrote:
>On July 29, 2004 at 09:35:57, Daniel Shawul wrote:
>
>>On July 29, 2004 at 09:21:10, Uri Blass wrote:
>>
>>>On July 29, 2004 at 09:05:50, Jorge Pichard wrote:
>>>
>>>>[Event "Arena tournament"]
>>>>[Site "S0025128408"]
>>>>[Date "2004.07.29"]
>>>>[Round "2"]
>>>>[White "Baron140b2"]
>>>>[Black "Movei00_8_251s"]
>>>>[Result "1-0"]
>>>>[ECO "A00"]
>>>>[Opening "Clemenz (Mead-Basman-de Klerk-Welling) Opening"]
>>>>[Time "00:11:54"]
>>>>[TimeControl "1800+0"]
>>>>[Termination "normal"]
>>>>[PlyCount "121"]
>>>>[WhiteType "program"]
>>>>[BlackType "program"]
>>>>
>>>>1.h3 e5 {+0.23/11 28} 2.a3 Nf6 {+0.55/11 35}
>>>
>>>
>>>This game suggests that opening book is not very important
>>>Baron starts 1.h3 2.a3 and still wins.
>>>
>>>learning can be more important and baron could beat movei more convinicingly by
>>>repeating this line again and again.
>>>
>>>Dann Corbit's book has no reply to 1.h3 so movei cannot learn.
>> why don't you add the line if it is not in the book.
>> danchess learning add played lines if not in the book and
>> change weight if it is.
>> daniel
>
>If the program lost the game then what do you add when you do not want to repeat
>a lost game?
>
>Note that in this case I think that adding the first moves of the game could
>help because in that case the program could have more time later and not blunder
>but I do not like it as a general solution.
>
>Uri
>Uri
Of course e5 wasn't a losing move, so you can safely add it. In Averno, I add
new moves even when I lose. I save search time along with the move, eval and a
penalty for losing the game. I have a threshold that the program uses to avoid
the move. In this case, I add e5, with a learnt value that is a combination of
the eval returned by search and the penalty for losing.
Next game, after h3, I check expected time to think. If lower or equal, I then
check if the learned value is above the threshold. If so, I play the move
instantly. If I later lose, I add another penaly. If I keep losing, at some
point the number of penalties would be big enough not to play the move. Then,
I'll add the move to the hash table with a very bad score and a high draft (I
also save the draft in the learning part of the book), and start searching. The
program will choose a different move without needing to search that bad move.
I do this for all the moves along the opening. I have definible parameters for
number of moves to add, threashold, randomness, etc.
My learning works in the long run, it is not an "agressive learning". It will
repeat losing moves some times before discarding them, but eventually it'll play
only sound moves.
José C.
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.