ZirconiumX wrote:There are some programs which research after LMR if value >= beta, some if value > alpha, and some don't research at all.
FF currently uses value >= beta, but I haven't tested this. This is why I ask if anyone else *has*.
Something tells me that value > alpha is best, because it allows Cut-Nodes AND PV-Nodes, rather than either only Cut-Nodes (value >= beta), or nothing (no research).
Matthew:out
Thanks Matt. You're right!!
I replaced >= beta by > alpha and DiscoCheck scored over 54% after 2000 games against itself (easily significant).
Bob: yes there are much fewer PV nodes, but triggering a full window research in those realatively fewer nodes is much costlier, so overall the difference in node count is actually quite significant (ie. >alpha is slower in time to depth, but more accurate, in the end tha accuracy wins for me).
One case where >= beta makes sense is if your engine doesnt reduce at PV nodes, of course (in which case >alpha and >= beta will produce the same result).
You should not draw conclusions on a few test positions, rather over complete games, where the difference is not so big. >alpha is still the right answer, because any score > alpha when you are expecting all moves to fail low indicates that something is not right. And most likely it will still fail low on the re-search when the depth returns to normal, and that is a LOT cheaper than accepting that fail-high (or actual score from a search that is too shallow) and then triggering a window-widening search, only to discover it really doesn't fail high like you thought...
IE it is better to just verify that the fail-high move is really going to fail high, rather than having to search ALL moves at this ply to see if that single move will fail high normally...
Bob,
Yes, I know that you prefer a gauntlet of engines over games against itself, but please, where did Lucas mention test positions?!?!
Lucas,
Glad to help, that's 2 improvements I've made to DiscoCheck. According to Sigma Chess, that means you've got about +30 ELO because of that.
Don,
That is an interesting idea. Maybe you could have a variable margin over depth, the idea being that at a higher depth, you may miss more. But then you pay the price of speed. Maybe that is something for you to try, Lucas? (*hint*)
Matthew:out
I didn't say "test positions". I said "positions". 2000 games is a marginal test with a significant error bar. Probably too big to measure which of the two is better.
>alpha is the only thing that makes any sense to me, otherwise you invite errors and search instability into the equation... You don't want to accept value > alpha unless it is with a proper search to the correct depth...
bob wrote:
I didn't say "test positions". I said "positions". 2000 games is a marginal test with a significant error bar. Probably too big to measure which of the two is better.
Wrong!
Of course if the score is 50.1% after 200 games, it's not significant, but the 54.5% I got *is* significant. Just do tha math to convince yourself !
bob wrote:
I didn't say "test positions". I said "positions". 2000 games is a marginal test with a significant error bar. Probably too big to measure which of the two is better.
Wrong!
Of course if the score is 50.1% after 200 games, it's not significant, but the 54.5% I got *is* significant. Just do tha math to convince yourself !
bob wrote:
I didn't say "test positions". I said "positions". 2000 games is a marginal test with a significant error bar. Probably too big to measure which of the two is better.
Wrong!
Of course if the score is 50.1% after 2000 games, it's not significant, but the 54.5% I got *is* significant. Just do the math to convince yourself!
bob wrote:
I didn't say "test positions". I said "positions". 2000 games is a marginal test with a significant error bar. Probably too big to measure which of the two is better.
Wrong!
Of course if the score is 50.1% after 2000 games, it's not significant, but the 54.5% I got *is* significant. Just do the math to convince yourself!
We need Jesus Munoz's magic tool for this.
Matthew:out
no need, a few very elementary calculations: empirical mean and stdev and gaussian distribution function. i do that in a spreadsheet in less than 30 sec
bob wrote:
I didn't say "test positions". I said "positions". 2000 games is a marginal test with a significant error bar. Probably too big to measure which of the two is better.
Wrong!
Of course if the score is 50.1% after 2000 games, it's not significant, but the 54.5% I got *is* significant. Just do the math to convince yourself!
We need Jesus Munoz's magic tool for this.
Matthew:out
no need, a few very elementary calculations: empirical mean and stdev and gaussian distribution function. i do that in a spreadsheet in less than 30 sec
bob wrote:
I didn't say "test positions". I said "positions". 2000 games is a marginal test with a significant error bar. Probably too big to measure which of the two is better.
Wrong!
Of course if the score is 50.1% after 200 games, it's not significant, but the 54.5% I got *is* significant. Just do tha math to convince yourself !
I am not talking about the result of the change. I was talking about your statement about affecting the size of the tree due to researches and where they are done. For me, first rule of investigating a change is to understand how it affects the tree. If you play very fast games, and you are changing the shape of the tree significantly, the results can easily be skewed for, or against the change, where the results might be completely different at a longer time control. In fast games, reduced searches often drop right into q-search, which is cheap. In long games, they don't...