Root LMR idea

[jd1]

Hi,

I had an idea on how to LMR at root inspired by the "easy move" thread, and would be very interested to know your thoughts.

What I do, is I search each root move until and including depth 4 with an infinite window. The root move list is sorted, and I record how many moves are -50 (or some other constant) centipawns worse than the best move. Then in future iterations I reduce all such moves, excluding extended moves and tactical moves. If a move fails high it is researched to the full depth and I do not reduce in it the next iteration.

Test results (at 10' + 0.2') are looking really good (+30 elo) after 700 games but of course that is not nearly enough games. And I doubt LMR at root can be worth that much.

Jerry

[jdart]

depth 4 is a shallow search so the idea that these moves are good to reduce later on the search is a bit dubious, IMO.

I have used the node count for root moves as a means to find reduction candidates. If the node count is relatively low that probably means the move has seldom if ever been the PV move, and it may have also have a quick refutation. But because the node counts are cumulative I am using information that continues to get better as the search progresses.

--Jon

[hgm]

There is no need to use an open window. You could just shift the null window doen 50 cP, and if you fail high first do a re-search with an unshifted null window before re-searching with an open window.

[lucasart]

Why would you not do LMR at the root anyway ? I treat the root node exactly like any other PV node (including for search reductions).

[jd1]

Why would you not do LMR at the root anyway ? I treat the root node exactly like any other PV node (including for search reductions).

Thanks to all for the helpful comments.

Lucas, I tried standard LMR at root but it didn't work for me. Maybe I didn't play enough games however.

Jon, you are absolutely right - I played longer games 10' + 0.5' and the result was clearly negative.

Jerry

[diep]

Hi,

I had an idea on how to LMR at root inspired by the "easy move" thread, and would be very interested to know your thoughts.

What I do, is I search each root move until and including depth 4 with an infinite window. The root move list is sorted, and I record how many moves are -50 (or some other constant) centipawns worse than the best move. Then in future iterations I reduce all such moves, excluding extended moves and tactical moves. If a move fails high it is researched to the full depth and I do not reduce in it the next iteration.

Test results (at 10' + 0.2') are looking really good (+30 elo) after 700 games but of course that is not nearly enough games. And I doubt LMR at root can be worth that much.

Jerry

hi,

just hopped on quickly through CCC, so missed the threads you refer to (no time to read those); the +30 elo is not because of the LMR, yet because of sorting the rootmoves pretty ok.