Well, but then I could use an even higher number, as long as I was sure the moves pruned where all moves with zero chance of producing a beta cut-off. Which you can never be 100% sure except by a very deep search, but there just is a limit to the amount of moves/positions you can search. But it is just an experiment and as you can see in the futility margins I use (post above), they are growing at a rate that is on the order of the square of the normal rate for Stockfish futility margins so that should undercut most oversights. And also the reductions are no absolute numbers so you could rephrase it a bit by thinking; "Are there moves that need to be searched 24 plies deeper than other moves, before you can be certain whether they are good or not?". Then I think the answer is yes so Larry I am not yet ready to accept that a search difference of 24 plies is silly !lkaufman wrote:I don't quite follow where the number 24 plies comes from, but to my thinking ANY algorithm that applies to the last 24 plies is silly. Anything that is valid anywhere near that far from the leaf almost certainly must be valid at any distance from the leaf.
True I suppose, but if you cut back a 24 ply tree with ten moves that still leaves a fourteen ply deep tree, that is still a significant save if you could be sure at the root there is a completely irrelevant position. So I'm not completely buying your argument Larry. That's why I said at a certain point, smart algorithms will have to stop searching and store the search-result permanently in position learning, as long as the search-result is deep enough below alpha. In order to do that you can't use a null window search centered on alpha, you will have to go below it, but especially in a cluster I think you will be better off leaving alpha beta anyway. That was my argument.Furthermore, it is almost certainly wrong to prune ANY move (except for something like bishop underpromotion that is extremely rare) more than around ten moves back from the leaf, because pruning at ten moves back eliminates 99.9% of the underlying tree assuming a branching factor of 2; how much more can you save by pruning even earlier?.
Can you clarify exactly what you are pruning or reducing 24 plies back, but not 25 plies back?
The pruning rules are not changed much from Stockfish here and neither is the static evaluation much different so that works the same as in Stockfish futility pruning. Only the futility margins after a while are growing much faster than they do near the leaves. I'm not saying it works better this way but it is also not completely useless. I'm just testing what kind of holes this leaves in a search, compared to a more normal search.
Eelco