Author: Robert Hyatt
Date: 17:16:58 06/18/02
Go up one level in this thread
On June 18, 2002 at 18:43:29, Sune Fischer wrote:
>On June 18, 2002 at 17:56:58, Robert Hyatt wrote:
>
>>On June 18, 2002 at 16:02:45, Sune Fischer wrote:
>>
>>>On June 18, 2002 at 12:11:57, Robert Hyatt wrote:
>>>>
>>>>OK... define "MATE" as 32768, for now.
>>>>
>>>>If you find a move at the root that leads to a mate in 4, the score
>>>>will typically be 32768 - 8 since you find you are mated at ply=8 and
>>>>you return MATE-ply (I do, anyway). Therefore, the root score that
>>>>gets backed up will be 32760, and so far, all is well. That means that
>>>>the "mate score" that gets returned from ply = 8 has to be - ( MATE - PLY )
>>>>so that - is bad for the side on move at ply=8.
>>>>
>>>>Note that you have 4 moves for the program in the path, the move at
>>>>ply=1, 3, 5 and 7. When you store a hash table entry at ply=7, you
>>>>should store MATE-1, since the move at ply 7 will result in a MATE at
>>>>ply=8. At ply 5, you should store MATE-3, at ply 3 you should store MATE-5
>>>>and if you want to store the entry at the root, it would be MATE-7.
>>>>
>>>>All this means is that at any point in the path that terminates in a mate
>>>>score of any kind, you have to adjust the score so that it is a mate-in-N
>>>>from the current ply, not from the root. Because the root score is wrong
>>>>at the current ply, since we are _closer_ to the final mate than we were at
>>>>the root of the search.
>>>>
>>>>If that isn't clear enough, ask again. You can find how I return a mate score
>>>>by looking at the bottom of search.c in Crafty (look for string MATE). You can
>>>>find how I adjust the mate score before I store it in the hash table by looking
>>>>at hash.c, module Store()...
>>>>
>>>>Bob
>>>
>>>I think that _is_ what I'm doing!?
>>>
>>>#define CHECKMATE 100000
>>>#define CMSCORE 99000
>>>
>>>search() {
>>>..
>>> if (!movesfound) {
>>> if (InCheck(wtm))
>>> score=-CHECKMATE+depthroot;
>>> else
>>> score=DRAWSCORE;
>>> }
>>>...
>>>}
>>>
>>>hasprobe() {
>>>...
>>> switch (phash->flag) {
>>> case HASH_EXACT:
>>> if (abs(score)>CMSCORE){
>>> if (score>0)
>>> score-=depthroot;
>>> else
>>> score+=depthroot;
>>> }
>>>...
>>>}
>>>
>>>hashstore() {
>>>...
>>> if (abs(score)>CMSCORE){
>>> if (score>0)
>>> score+=depthroot;
>>> else
>>> score-=depthroot;
>>> }
>>>...
>>>}
>>>
>>>"depthroot" is just distance to the root position we are searching, I assume it
>>>is identical to yours "ply" variable.
>>>
>>>Something must be wrong, I get weird results, it does work fine, it seems, but
>>>when I look at the log I'm worried.
>>
>>One last thing: when you store the score, how do you do that? IE I use N
>>bits for the score, and to avoid sign-extension for negative numbers, I add a
>>constant larger than any possible negative score and then store the sum. If
>>you do that, you aren't suffering from an overflow??? IE 1 bit can make a big
>>difference.
>
>No overflow, I'm using a full 4 byte signed integer for the score.
>That's because I'm using milipawns, so two bytes is not enough.
>Three bytes would be enough, but that's just an optimization I can play with
>later.
>
>>>
>>>This position was from a quick test game, it's mate in 3, but I get mate in 1
>>>three times in a row, but it doesn't have a problem playing the right moves,
>>>really strange...
>>>[D]1r2qrk1/n5p1/1p1p2P1/p2p2P1/2P1pPK1/P1PB2P1/2Q5/1R5R w - - 0 1
>>>
>>>-S.
>>
>>
>>I get a mate in 3. I make a move, get mate in 2, make a move and get
>>mate. All from the hash table after the mate in 3 is found...
>
>...but you couldn't find anything wrong with what I pasted?
>I don't see anything wrong with it either, matescore is first adjusted for
>distance to root, then when recorded to hash it is readjusted to be the distance
>from where I save to the mate, and when I probe it's adjusted again by adding
>the current distance to root.
>Why tha' heck isn't that working??
>
>-S.
I will look at it carefully tonight. One thing I did find years ago was that
adjusting everything that is >= MATE-100 was a mistake. IE I don't adjust
"bounds", just EXACT scores. That used to cause some interesting mate scores
to pop up from time to time...
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.