late move reduction

[nionita]

Last september I introduced in Barbarossa variable LMR, i.e. there is a LMR reduction level which vary depending on number of re-searches: when more re-searches are done, level goes down (which means: less/later reductions), when less re-searches - level goes up, which means more/earlier reductions.

The implementation is a bit more complicated, but this is the idea.

This brought a few ELO points more. The change is already in the release 0.3.0.

[flok]

Oh and I should also do the reduction after 3-4 moves and not the first one right away

[flok]

One further question: do you need to have evaluated 3-4 subtrees? I mean do you need to have the score of 3-4 search()es before you can reduce?
Because it is possible that in some nodes I started, say, 4 threads so the 5th move will be evaluated with the value with which the node started.

[hgm]

In micro-Max only hash move, captures, and Pawn moves are exempted from reduction. So if there are no captures and all Pawns are blocked, only a single move is searched before starting to reduce. LMR is applied in all nodes with d >= 4 (reducing the search of the reply from 3 to 2).

[jdart]

You can also try a two-stage re-research. First, re-search with no reduction but still using a zero-width window. Then if that still fails high, re-search with no reduction and high bound of beta. Some engines do the two stage thing but only if the reduction was above a certain amount.

There are lots of other knobs. I exclude killer moves from LMR but I think Stockfish recently put in a change that either allows reduction of all killer moves or maybe after the 1st one.

--Jon