It is not the same at all.Aaron Becker wrote:Isn't this analogous to comparing the score of two moves searched with nominally equal depth when the extensions and reductions applied to each move are not the same? In any node which doesn't experience a cutoff, when we apply LMR we are implicitly committing ourselves to comparing searches with unequal depths just by looking at the resulting scores. If you're arguing that this is an unprincipled way of searching, then formally I agree, but in practice the results are compelling despite the lack of a rigorous theory to back the technique.bob wrote: I'd be happy to show you positions where A(16) has a higher score than B(17), yet when you stop and choose A and then ponder it fails to below the score for B. Happens every time you flip-flop on two successive iterations.
Do what you want, but don't try to convince me it is correct.
If there is a fundamental difference between GCP's suggested approach and score comparisons between reduced and non-reduced nodes, could you clarify what that difference is? I think that reductions are safer because we only make them when there is reason to believe that the move being reduced is not best, but in principle they seem very similar to me.
In normal search, you reduce that which you do not expect to do well. In this form of search, you are radomly searching some moves deeper and some shallower, and then comparing the scores. you will note that you do not play a PV where reductions were done. At least in Crafty's PV the moves shown will add up to the proper number of plies. That's a different issue. Some moves are simply easier to search, and you can go deeper if you choose to spend equal effort on each move. But why spend the same amount of effort on easy to search/dismiss moves as you spend on the actual move you are going to end up playing???
This unsynchronized stuff is bogus from many directions. Comparing the scores/depths is problematic. Searching two moves independently and unsynchronized is inefficient because it seriously degrades the alpha/beta process.