I'm experimenting with late move reductions in Anatoli and I don't see a clear difference between reducing after the first legal move (if certain conditions are met, don't reduce captures etc) and reducing after the third or fourth move or so.
In the middle game this is probably because the moves immediately after the first are often captures that aren't reduced anyway. I didn't try many endgame positions yet, perhaps it does make a difference there.
What are the experiences of others with this?
I also get the impression from my (albeit limited) testing that the positive effect of LMR (either after the first legal move or later) becomes less and less at longer time controls. Could it be that bullet and blitz tests are somewhat misleading and that the effect becomes zero or even negative at long time controls? Sort of diminishing return?
Martin wrote:I'm experimenting with late move reductions in Anatoli and I don't see a clear difference between reducing after the first legal move (if certain conditions are met, don't reduce captures etc) and reducing after the third or fourth move or so.
In the middle game this is probably because the moves immediately after the first are often captures that aren't reduced anyway. I didn't try many endgame positions yet, perhaps it does make a difference there.
What are the experiences of others with this?
I also get the impression from my (albeit limited) testing that the positive effect of LMR (either after the first legal move or later) becomes less and less at longer time controls. Could it be that bullet and blitz tests are somewhat misleading and that the effect becomes zero or even negative at long time controls? Sort of diminishing return?
Hi Martin,
I have actually thrown out any history dependency and delegate (fractional) reduction to the expected cut child before I actually probe the hash with an adjusted depth. Beside the number of the quite moves already tried at the expected parent all node, I consider properties and influence of the move after updating the board and processing some informations (attack- and defend sets, passers etc.) to have a better static knowledge-base to determine whether and how much I extend or reduce. Likely the first three quite moves are never reduced and I reduce potentially the more, the later the move is (> 6, > 12) and whether I already tried a move with that piece before. Of course move ordering is crucial.
But I am not a painstaking tester and don't want to become a slave of test-procedures to play thousands if not millions of games to maintain some local maximum
Martin wrote:I'm experimenting with late move reductions in Anatoli and I don't see a clear difference between reducing after the first legal move (if certain conditions are met, don't reduce captures etc) and reducing after the third or fourth move or so.
In the middle game this is probably because the moves immediately after the first are often captures that aren't reduced anyway. I didn't try many endgame positions yet, perhaps it does make a difference there.
What are the experiences of others with this?
I also get the impression from my (albeit limited) testing that the positive effect of LMR (either after the first legal move or later) becomes less and less at longer time controls. Could it be that bullet and blitz tests are somewhat misleading and that the effect becomes zero or even negative at long time controls? Sort of diminishing return?
I think that the effect of LMR is dependent on the implementation so there is no general answer for your question.
different programmers have different rules when to use LMR.
Martin wrote:I'm experimenting with late move reductions in Anatoli and I don't see a clear difference between reducing after the first legal move (if certain conditions are met, don't reduce captures etc) and reducing after the third or fourth move or so.
Of course there is no difference since you are reducing an ALL node. If you are there then it is not a CUT node. If there is a difference then you are reducing a possible cut node.
In the middle game this is probably because the moves immediately after the first are often captures that aren't reduced anyway. I didn't try many endgame positions yet, perhaps it does make a difference there.
What are the experiences of others with this?
First moves are 'good' captures so you have immediatly a cut off. So you can start to reduce after the first move as long as you don't reduce those good captures or any move flagged as good.
Endgames is about deep search, so you'd better not reduce too much there or you won't see any pawn promotion.
I also get the impression from my (albeit limited) testing that the positive effect of LMR (either after the first legal move or later) becomes less and less at longer time controls. Could it be that bullet and blitz tests are somewhat misleading and that the effect becomes zero or even negative at long time controls? Sort of diminishing return?
There is no diminishing return, you reduce/prune too much where you should not and you have plenty of re-search as a consequence of your incorrect search. Look at your re-search counter of lmr/new pv. Are you re-searching 1% of your reduced branch or something like 10% ?
Harald Johnsen wrote:
Of course there is no difference since you are reducing an ALL node
I guess you mean 'if' instead of 'since'? Then I agree, of course.
Harald Johnsen wrote:
There is no diminishing return ..
Without actually testing in games it's hard to say. But testing is timeconsuming of course, because you would need to play a lot of games at longer timecontrols to see if there is indeed a diminishing return (less and less positive effect of LMR with more time).
One question is, I think, how much LMR hurts positional play at longer time controls because it destabilizes the search. If the tactical gain becomes less (as far as I see now that is indeed what happens) while the positional damage remains you would expect that the overall gain becomes smaller and smaller as time increases.
I didn't use history counters, btw, only simple static criteria (don't reduce captures, checks, moves that are extended, etc). I guess history counters might help positionally.
Btw, I'm talking about gain in playing strength of course. The percentage of researches doesnt't tell me anything about that (unless it's 100%, in which case I have a bug ).
