LMR question

[Uri Blass]

Do you reduce late root moves in LMR

If I understand Glaurung's source correctly Glaurung does not reduce late root moves and I wonder if it was tried and found counter productive.


It is clear that there is not a lot to gain by reducing late root moves
because most of the time for root moves is used for the first moves but I wonder if there is a reason to be not consistent.

Note that I also not reduce root moves simply because I did not think about it and I have different functions for root moves and for normal alphabeta.

Uri

[bob]

I can do it easily. Testing was inconclusive for me, so I simply put it on the "to-do" list to resolve later... I am probably a bit more restrictive than Tord also, in that I don't do LMR until after I have tried all move ordering tricks including killer moves (2) and another 4 history moves as well. I plan on testing without that restriction also..

[Tord Romstad]

Do you reduce late root moves in LMR

If I understand Glaurung's source correctly Glaurung does not reduce late root moves and I wonder if it was tried and found counter productive.

Almost. It was tried briefly, with unconclusive results.

It is clear that there is not a lot to gain by reducing late root moves because most of the time for root moves is used for the first moves but I wonder if there is a reason to be not consistent.

Perhaps it isn't a very good reason, but this is how I think:

When we choose to reduce some move because it looks unpromising, we always take a risk. It is possible that the move will look bad in the reduced depth search, but would score well in a deeper search. This risk is always there, but it isn't equally important at all nodes. At most nodes in the search tree, an incorrect search decision is unlikely to propagate up to the root. At the root node, however, an incorrect pruning or reduction decision is guaranteed to produce an incorrect score and best move, more or less by definition. It therefore makes sense to be more conservative at the root than at internal nodes.

For similar reasons, it is natural to avoid reductions at PV nodes, where the score is much more likely to propagate up to the root than at non-PV nodes. In principle, I want to do this, but to my frustration Glaurung plays clearly worse when I disable LMR at PV nodes. I wish I understood why. Most other authors report the opposite.

Tord

[Pedro Castro]

I have reductions at PV nodes.

DanaSah plays well. I can simply check the difference with any test of positions, It solves any position well, overalls if it is tactical.