Michel wrote:Can you test the new version at
http://alpha.uhasselt.be/Research/Algeb ... _versions/
The file is called
Regards,
Michel
PS. I replaced the sem_ functions by my own functions.
Code: Select all
typedef struct {
int value;
pthread_mutex_t mutex;
pthread_cond_t cond;
} my_sem_t;
void my_sem_init(my_sem_t *sem, int value){
sem->value=value;
pthread_mutex_init(&(sem->mutex),NULL);
pthread_cond_init(&(sem->cond),NULL);
}
void my_sem_post(my_sem_t *sem){
pthread_mutex_lock(&(sem->mutex));
sem->value++;
pthread_cond_signal(&(sem->cond));
pthread_mutex_unlock(&(sem->mutex));
}
void my_sem_wait(my_sem_t *sem){
pthread_mutex_lock(&(sem->mutex));
while(sem->value==0){
pthread_cond_wait(&(sem->cond),&(sem->mutex));
}
sem->value--;
pthread_mutex_unlock(&(sem->mutex));
}
Michel, something to try, maybe I'm wrong but I think there might be an error in the original 1.4 Beta5c and earlier Fruit and Toga sources. Have not checked earlier versions though.
My reasoning is the following:
The History Research in full_search () is tried when value goes over beta but I think the assumption for 'Late Moves' or History Pruning is that these moves will not increase
alpha 
. So if the LMR/History reduced search goes over alpha this assumption fails and a re-search is called for.
Actually if the reduced search goes over beta even for a 'Late Move' the/a previous move of the opposing colour is probably bad and a re-search (or a search of an early, better move) would to some likelyhood still be larger than beta, the value in that case is not really that important anymore
But a re-search still prudent in case the re-searched move falls into the alpha beta window after all.
If you do more than one ply of reductions, by Extended History Pruning or something to that effect probably the re-search should be done in steps not just a single ply extension, I don't see that in the sources at the moment but I have not tried a correction of this yet. >= alpha I just tried and I think it might be an improvement? But very early days
! I just thought this looks wrong in the code but on the other hand many people must have studied this part of the Fruit and Toga sources
Line 783 in search_full.cpp
Code: Select all
// history-pruning re-search
if (HistoryReSearch && reduced && value >= beta) { // [EdG: value >= alpha seems a correct bugfix here?]
ASSERT(node_type!=NodePV);
new_depth++;
ASSERT(new_depth==depth-1);
value = -full_search(board,-beta,-alpha,new_depth,height+1,new_pv,NODE_OPP(node_type),cap_extended,ThreadId);
}
move_undo(board,move,undo);
Regards, Eelco
