Interpretation of LMR

[brtzsnr]

I know that LMR was recently discussed in a couple of threads, but I'd like to propose an alternative interpretation of search with LMR.

Let's suppose that LMR is implemented as: reduce search depth of all quiet move by one. A quiet move means is not promotion, capture, check, killer, hash move, etc.

Now an arbitrary path on the search tree has 1 * V + 2 * Q + QS moves where V is the number of violent moves, Q is the number of quiet moves and QS is the number of moves in the quiescence search (all violent). If quiescence search is unbounded then a search with LMR is a search in which all paths contain at most M quiet moves.

Does this makes sense? Has anyone tried limiting search depth by the number of quiet moves instead of number of all moves?

[Angrim]

I think you would run into a lot of positions where that search would explode badly. Even unbounded qsearch sometimes becomes a problem, adding in checks, promotions, and killers to the list of moves that are always extended would make for remarkable tree growth.

[hgm]

This is equivalent to extending all captures by half a ply, after renormalizing what 'ply' means. The problem is that many captures are very bad, or necessary (but trivial) refutations to non-captures that were very bad. And you would extend all those bad moves, so that the search focusses on nonsense branchess in positions where quiet moves are what the doctor orders.

There also is an inconsistency in what you propose: a reduced move that fails high would in fact become a hash move, except that it was discovered only after the fact. (Perhaps because the hash entry was overwritten.) So it would be inconsistent to keep the reduction, if you don't reduce hash moves.

[Dann Corbit]

I know that LMR was recently discussed in a couple of threads, but I'd like to propose an alternative interpretation of search with LMR.

Let's suppose that LMR is implemented as: reduce search depth of all quiet move by one. A quiet move means is not promotion, capture, check, killer, hash move, etc.

Now an arbitrary path on the search tree has 1 * V + 2 * Q + QS moves where V is the number of violent moves, Q is the number of quiet moves and QS is the number of moves in the quiescence search (all violent). If quiescence search is unbounded then a search with LMR is a search in which all paths contain at most M quiet moves.

Does this makes sense? Has anyone tried limiting search depth by the number of quiet moves instead of number of all moves?

Do a search through PGN of games played by high-end opponents.
Most good moves are quiet moves.

There is nothing about a quiet move that makes it less important than a capture.

The objective of pruning should be to spend less time on moves that are not promising (which is what null move, LMR and the like accomplish). I do not think your plan will accomplish that.

[hgm]

There is nothing about a quiet move that makes it less important than a capture.

Captures are more forcing, because you have to recapture or lose material. So they are ideal delaying tactics to push threats over the horizon, which is why evaluating them in general requires more depth than for non-captures. (Except that some non-captures also are threats that force an evasion, and those deserve deeper search as well. This is the 'tacticall connected' stuff of Glaurung's LMR.)

[Dann Corbit]

There is nothing about a quiet move that makes it less important than a capture.

Captures are more forcing, because you have to recapture or lose material. So they are ideal delaying tactics to push threats over the horizon, which is why evaluating them in general requires more depth than for non-captures. (Except that some non-captures also are threats that force an evasion, and those deserve deeper search as well. This is the 'tacticall connected' stuff of Glaurung's LMR.)

Yes, but qsearch takes care of that.

[hgm]

No. In Q-search you can stand pat to turn a blind eye to impending doom. E.g. if a Knight is forking my two Rooks, at d=2 a senseless non-capture would notice that, because the opponent cashes the Rook in the second ply. But if I throw a d=2 search on a capture QxQ, the NxR would make me stand pat Q-vs-R ahead, while the PxQ recapture would make me stand pat with an even score, even though I am still going to lose the exchange.

[Engin]

i tried also reduce of bad captures with a slightly success, because most of the bad captures with SEE < 0 are not comes near to alpha.

[cdani]

i tried also reduce of bad captures with a slightly success, because most of the bad captures with SEE < 0 are not comes near to alpha.

In Andscacs I reduce them with one when depth > 1.

[Engin]

can be reduce even more i guess