Author: Robert Hyatt
Date: 12:03:53 01/10/98
Go up one level in this thread
On January 10, 1998 at 14:53:44, Don Dailey wrote: >>I don't know of anyone who will make a statement of the form: "I/We do >>singular extension, it helped us a lot, we proved to ourselves that it >>is a good thing, and for you guys who are curious, here is a position >>that we think it helps a lot in, and here is some other cool info >>involving lots of other interesting numbers, for instance, look how many >>more we get in the ECM suite with this!" > >>If someone wants to make all or part of the above statement, I would be >>delighted to hear it. > > >I cannot make this statement either. I implemented it a while back and >it seemed to hurt the program slightly. But I'm not going to state >the contrary either. There are too many factors and implementation >details that could make a difference. > >But I'm not really too optimistic about these extensions because I don't >think singularity is the key point although I do believe it is >definitely >relevant. > >The key point in my mind is to find a class of extensions that will make >you understand things better at the end node. I tend to think in terms >of "getting out there deeper" but I think that is wrong, the real issue >is getting a more accurate score. I admit the "easiest" way to do this >is to 'get out there deeper' but extensions always have the side effect >of decreasing your minimum depth somewhere else, so we need to zero in >on which moves these are. Singular moves are probably nothing more >than >"likely candidates", but still with low probability of helping. > >I'm guessing some variation of singularity that is more "picky" might >be a better choice. I may re-implement sigular extensions and >do some more experimentation. > >- Don My only comment about these extensions, after the obvious "I tried them years ago but thought they were too expensive" is that both of us have been absolutely destroyed by them. IE in every game I have played against DT and its successors, it found incredibly deep (winning) lines that we followed having no idea that we were even in trouble. I posted one such position here or in r.g.c.c a while back where they found a forced way to win a bishop and *saw* they were winning, and this conbination ended up 30+ plies deep, with none of the moves being captures or checks. We saw the problem 10 full moves after they first failed high. I also recall one of the last two ACM events where *Socrates was actually searching more NPS, and searching a ply or two deeper (claimed iterative depth) but was totally crushed by a long combination that they saw again at least 5-10 moves before *Socrates saw it. Bottom line is this is complicated to implement, certainly won't be done in 3 months or so, because there are *so* many side-issues they explained in their ICCA paper, including the sticky transposition table, plus other things needed to make this work. I don't like the idea of "singularity" either. However, their results do speak for themselves. And don't forget Genius is using a form of singular extensions (although I believe he uses a much weaker form). My first cut at these in 1978 involved many months of testing and tuning. And it ended up costing me a ply to find deeper tactics. Unfortunately it mean dropping back from 5 to 4, which was too much. I'd probably give a ply today if I could follow some of the deep forcing lines I see from time to time. It is still on my list to re-visit. But it has to wait until I have a lot of time because it is a complex algorithm to implement, if the whole thing is done and done right...
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.