TalkChess.com

I have some problems with LMR in late endgames, especially pawn endgames (but also minor piece endgames). Scores become unstable and winning lines are seen much (8 plies) or so later. This with a reduction of 1. Especially the reduction on PV nodes seems problematic. Is this normal? Should I disable LMR in late endgames?

metax wrote:I have some problems with LMR in late endgames, especially pawn endgames (but also minor piece endgames). Scores become unstable and winning lines are seen much (8 plies) or so later. This with a reduction of 1. Especially the reduction on PV nodes seems problematic. Is this normal? Should I disable LMR in late endgames?

I did that (disabling) in pawn endgames and if one of the kings is alone due to similar reasons. Actually in my case LMR was more like a weight for the search set instead of relieve.

Technigues like LMR and null-move are counterproductive in Pawn endings.

What can be said in general about LMR in endings in general (not pawn endings)? Are any programs known to have different rules for LMR in the endgame in general than in the middlegame? If so do they look at more moves or less moves before reducing? On the one hand, you might want to reduce less often in the endgame, but in general there are less legal moves to start with, so it's not clear.

lkaufman wrote:What can be said in general about LMR in endings in general (not pawn endings)? Are any programs known to have different rules for LMR in the endgame in general than in the middlegame? If so do they look at more moves or less moves before reducing? On the one hand, you might want to reduce less often in the endgame, but in general there are less legal moves to start with, so it's not clear.

I don't know about others, but I do LMR everywhere but pawn endgames and 'king_alone' positions.The problems arise when the only way for the program to collect knowledge remains the deep search. Let me explain. If one don't use anything special for move ordering in the late endgame,and the evaluation contains no reliable code for passers, no reliable distance to passer measurements, no fast pawn enclosure lookup, e.t.c. , the only way to get there is to search as deep as it's possible. If the engine is using very aggressive reductions, some of the suitable variants will be reduced and therefore, searched again on an new iteration or even during new search. As a result, proper variant could be seen slowly (if at all). This is just my impression and i would like to read others opinions on that subject!

In my case, I don't allow reductions in neither PV nodes nor passed pawn advances. This help a lot.