Author: Robert Hyatt
Date: 21:42:38 08/08/01
Go up one level in this thread
On August 08, 2001 at 15:10:16, Bruce Moreland wrote: >On August 08, 2001 at 14:03:50, Frank Phillips wrote: > >>On August 07, 2001 at 16:03:34, Bruce Moreland wrote: >> >>>I've often heard people state that null move with R=3 is better than with R=2, >>>but I have never ever ever gotten a test result that indicates this. >>> >>>I've tried everything. I've tried it throughout the tree, I've tried it near >>>the root, and I've tried it near the tips. >>> >>>My measurement standard is ECM positions solved, which *always* goes down. >>> >>>What are other people doing that I'm not doing, or are people testing in some >>>other way, if so is their way better or worse? >>> >>>I would test Crafty both ways (it's currently doing R=3 some places), but my >>>machines will be busy until after the WMCCC. >>> >>>bruce >> >>What makes R=1 better than R=0. >>What makes R=2 better than R=1. > >2 is a magic number. It works better than 1 and better than 3 in my experience. > It also works better than 2.5 and 1.5, on planet Bruce. > >>What makes R=n+1 better than R=n. >> >>I got no particular benefit I can identify from Heinz's adaptive null move >>(R=2/3). But then again I have no reliable test methodology - other than the >>size of the tree, which is a bit smaller, but whether this is the correct metric >>I do not know, since not searching at all makes the tree smallest. > >Ernst is trying to do it right. I can pick on him in some cases, but I would >accept his results for this case, meaning, based upon the numbers he reports I >would use it too. He goes faster and he doesn't lose anything. > >Of course, he may have gotten lucky. ECM is a big part of his test. I have >enough experience with ECM to know that you can make a minor change to your eval >function or soemthing and suddenly your ECM results go into the toilet. > >My own problem might be that I search for a local optimum. I do ECM and on my >quad and maybe I get 700. I make a tweak and I get 688. It's hard to take the >688 result without good reason, and it's hard to have a good reason. If I do 50 >games of autoplay with Crafty I end up with about the same score so there's not >a glimmer of significance in that. > >I have other tools for looking at ECM results. What if I get 12 fewer, but >those that I do get I get faster in 2/3 of the cases, and in general the thing >returns its PV's a little faster in 3/4 of the cases. How do I interpret >*that*? > >Especially if this all came about because all I did was I scale my bonus for a >strong-point knight depending upon which file the knight is on. > >This makes me doubt anyone's result for anything. > >The reason I'm asking about this R=3 thing is that I knew Vincent did pure R=3, >and he told me that "everyone" was doing it, which someone else now says means >only Vincent. Then I see that Yace is doing this Ernst thing, and so is Crafty. > I want to try to figure out if this has just become unquestioned conventional >wisdom, or if there's really something to it. > >It sounds like Bob is doing it because he likes the extra depth, and hasn't >taken into account any tactical failures. That may be the right way to evaluate >this, and it may not be, but it's good to know, since I'm *only* concerning >myself with tactical failures. > >bruce > I should have added that I played this a good while before Ernst reported his results in the JICCA. Stanback started me on this trail about 4 years ago or so. He suggested using R=2 normally, but using R=1 near the tips to avoid hiding the tactical problems that R=2 causes near to the tips (it collapses me right into the q-search, which is a problem). That got me to thinking at the time that R=2 to R=1 was not a bad idea, but the R=1 really hurt performance as you might guess. As my search got deeper (going from the P5/133 to the P6/200 was quite a jump in speed, roughly 2.5X or so) R=2 became more trustworthy and that was when I started toying around with R=2 or R=3, right after the match you and I played for fun one night where you used R=3 and we didn't see anything bad happen to Ferret as a result. I believe that Ernst and I slightly disagreed on where to make the switch from 3 to 2... I chose my "shift point" after playing lots of test matches, while Ernst used test position results to adjust his. We ended up within a ply of each other if I recall, although I think he might have shifted further away from my shit point later on after more testing. I tried lots of things, including a gradual reduction from 3 to 2, using fractional ply reductions... and I even tried R=4 to 2. Which didn't look bad. However, I was a bit suspicious of R=3, so R=4 really got my super- stition active. I used to hate R=2 if you recall. And I am pretty sure that R=2 to R=3 is worse at blitz but better at longer time controls. But not _hugely_ better... just "better". But I used games to judge this, which might not be the best way. There is still plenty of things to try with this 'adaptive' behavior. I doubt the last word has been written yet... John Stanback served as the catalyst for me. It is possible that (a) he has experimented with this a lot as well, although I didn't hear him mention R=3/2 when he was talking about R=2/1 back then. (b) he heard about the adaptive R adjustment from someone else himself and was just relaying what he had heard rather than what he had tried. No idea whether he was the first to suggest this or not. In my case, it was simply a natural extension of his idea to go from 2/1 to test the 3/2 shift as well. I can certainly run some test suite of your choice with R=2 and R=3/2 to see the difference, if you want... >> >>Frank
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.