Author: Robert Hyatt
Date: 15:02:38 08/19/98
Go up one level in this thread
On August 19, 1998 at 15:16:59, Dan Homan wrote: >Yesterday I asked a question on ICC about extending simply >for being in the principle variation, the idea being to detect threats a >little earlier at the expense of possibly missing a good tactical shot. > >Bruce Moreland suggested that I test this with a tactical test >suite to get an idea for how much tactical zip I lose by doing this, and >to bin the results with respect to solution time to give additional >information. Below are the results of the test I ran. > >My implementation of the extension is to simply add some >fractional ply extension if we are in a pv node at depth = 1 or >depth = 0. I determine if this a node is a pv node by simply >comparing alpha and beta to the root values of alpha and >beta. > >For my program it turns out that if I use 3/4 ply fractional extension, >this adds 4 ply to a full pv line, and if I use a 2/3 ply extension, this >adds 2 ply to a full pv line. > >To test this idea, I used the subset of the ECM suite posted by >Jouni Uski a couple of days ago. The total problem set is 231 >rather challenging problems. I ran each problem for 60 seconds >on my Compaq Presario at home (it has a strange processor, but >as a practical matter is about 15% slower than a P133). > >Solutions were only counted as correct if they were held until the >end of the 60 seconds search period. > >Here is a table of the problem solution times for a 4 ply effective >extension in the pv, a 2 ply effective extension, and no extension. > > # of problems solved >time bin No pv 2 ply pv 4 ply pv >(seconds) extension extension extension > >0 - 15 30 30 25 >16-30 17 13 20 >31-45 13 13 12 >46-60 5 8 4 >--------------------------------------------------------------------------------------- > totals 65 64 61 > >So, as a general rule, the extension slows down the solution >times. In order to gauge this properly I would also need a good >"avoid move" testsuite to see if the pv extension would >substantially decrease the amount of time to see that a tempting >move is really bad. > > - Dan > >P.S. My game results on ICC seem no worse with the extension than >without it, but I am not sure that it is better. This has been done in two "flavors" over the past 30 years or so: #1 was suggested by no less than Richard Greenblatt and used in the mid- 60's... after computing a "new PV" (ie after backing a score up to the root) before saving the PV/score, first step back down to the end of the PV, and from that point, do a 2-ply search. If the score is worse, use the new score and see if it is better than the old "best". If not, toss it, if yes remember this as a new PV. If the two ply search "raises" the score, ignore the new value and use the value you computed for this PV with the normal search. Justification: extend the PV to see if you are doing something that is simply pushing something bad beyond the horizon. If so, the two ply search might reveal this with a lower score. But if the score goes "up" it is very speculative and can't be trusted. Results: I tried this in the early 70's myself, and found it was nothing more than a big way to waste time. #2 is the so-called PV extension. The question is, what is the justification for searching *just* the PV nodes to a deeper depth than the others? The answer is there isn't, and it is also simply a way to slow you down. It might find an occasional tactic quicker, but it is generally slower... My recommendation: toss this too... extending "just because" has to be bad...
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.