I always found the idea appealing. Unfortunately, I could never get good results.
There are few things that perhaps can improve NegaC
1. Instead of binary search use golden or fibonacci ratio
2. Instead of global MIN and MAX use some kind of aspiration windows
3. Instead of stop criteria based on score [while (min != max)...] stop when one move is better then others (this may lead to imprecise PV but possibly improve time to depth)
Today I failed with using null move verification search as a replacement for internal iterative deepening. The idea was to retrieve hash move after a verification search (in absence of a real hash move) and order it first. Of course it is stupid, because we know for sure that this move is rather suspect, as it didn't score above beta. If it did, we would have a cutoff and would not need any move ordering in first place
However, there just might be some point in updating killers, history and refutation moves if verification search produces a cutoff.
Inspired for the post about improving SEE, but not related to SEE, I tried a hash of captures, whose key is the combined keys of the source and destination of the pieces that can capture in a position, and the idea is to store the best capture.
After some tries I arrived to one version that is just level for fastest time controls, and possibly a little loss for 30 seconds or more. For the moment I dismissed it. May be I can use it or something similar for instead try to order better the captures.
My best attempt was by using it only alpha_beta, not in quiescence, and not store it if the best move was the first one.
I was fighting a lot trying to use somehow the threat move from the null move search.
At the end I inserted it as a killer move. After about 70k games the resulting gain was ~2cp, a bit smaller then the error margin (more engines involved in the games). But I let the code in place, 'cause I could not accept the idea that that information is useless.
nionita wrote:I was fighting a lot trying to use somehow the threat move from the null move search.
At the end I inserted it as a killer move. After about 70k games the resulting gain was ~2cp, a bit smaller then the error margin (more engines involved in the games). But I let the code in place, 'cause I could not accept the idea that that information is useless.
I have it also like a killer, with preference in move ordering and less reductions.
Stockfish had code that would extend a previously reduced move if the threat move from a null search was related to that previous move. They also disabled certain pruning/reductions for moves that appeared to counter the null move threat. However IIRC this code was removed from Stockfish at some point.
nionita wrote:I was fighting a lot trying to use somehow the threat move from the null move search.
Indeed, I also had that experience. The true threat seems to be masked most of the time by an unrelated harmless Q or R trade, which in the move ordering goes (quite justly) before cashing of an existing threat. The special evasion code that would use this information would then try to avoid that trade, rather than solving the real problem.
Basing the threat on SEE is much more reliable. If the node after nullmove uses SEE to order its captures, the capture with the highest SEE could be labeled as the threat move. The catch is that for efficiency reasons we often don't calculate SEE for good or equal captures, as they are ordered by victim value. LxH captures are easy to recognize, and always represent threats. But for the equal captures of protected pieces it depends on what follows. So you might be forced to subject them to SEE, at least in the node after null move.