TalkChess.com

[Eelco de Groot]

Sometimes I think Houdini does not do Nullmove, so it never stops searching... Recursive nullmove can shave off a lot of depth, no limit if it weren't for the reduction itself running out of enough plies to reduce.. After a reciprocal null move is done suddenly you have legal positions again. Very weird stuff, amazing that it works so well in general! I have no way of measuring it but I think, very uneducated guess, nullmove reduces very roughly about as much as Late Move Pruning and Futility Pruning (not counting LMR) in Rainbow Serpent, measured against non reduced moves. Just a gut feeling, could be very wrong. RS does LMP and Futility everywhere in the last 24 plies still, so compared to that Ivanhoe is like a knight with agoraphobia 10 plies, pffft. There must be a way to strengthen the verification search though. I know Ryan Benitez has discovered a trick to improve nullmove in the source in less than 5 minutes for Fruit for an improvement by twenty elopoints, and on this forum only Chris Conkie was told how to do it. But we never found out what it was. It can't be in new Fruit's verification search though because Ryan abandoned doing that.

Didn't sleep too well because I wanted to do something like that, how hard could it be to find such a thing? But I don't think I am really on to it yet. I did think of a way to 'link' nullmove and the verification search a little bit like in Glaurung was done by Tord with the threatmove that links nullmove search with the full search. But because the nullmove will have failed high in case of verification, you do not have a 'real' threatmove, as the response move to nullmove (I call this the 'repair' move) does not cross alpha. But on the plusside, you have tried almost every illegal move possible after the nullmove, trying to find that cut-off. On the downside again there is no way to know the best move in an All-node and the deeper you search the worse the estimate of the real value is going to be, if you even store such a thing. But still, I think there is something you could do with the stored move, if you do actually do store the move that is still called the bestmove in Stockfish, even if it is very little. I call the result a 'weak threatmove' and because verification search does not do a nullmove in the first ply, you can replace a 'real' threatmove with this 'weakThreatmove' that I thought up in the half hour I gave myself to find Ryan's five minute change. It is likely I messed up the implementation, but this is what I got done after half an hour between sleepcycles:

Previous version: