Author: William Penn
Date: 15:19:17 12/01/00
Go up one level in this thread
On December 01, 2000 at 17:07:41, John Merlino wrote: >On December 01, 2000 at 16:31:51, William Penn wrote: > >>Some observations re CM8000 user-defined personalities >> >>Regarding a CM8000 Selective Search setting of 32 (obtained by editing the >>appropriate byte in a personality file with a hex editor) versus the standard >>Selective Search maximum setting of 12: >> >>I used the equivalent of Shep's CM6666 and CM6666 Deep as the two personalities >>for my comparison tests. For purposes of easy reference here, I'll dub them >>CM8666 and CM8666 Deep. The CM8666 personality had Sel.Srch.=12. CM8666 Deep >>had Sel.Srch.=32, but was otherwise identical to CM8666. >> >>I used a typical middlegame position for my comparison tests, chosen partly >>because it didn't take too long to get to higher ply calculations. >> >>There was no difference in any Thinking Lines window information at 13 ply or >>less. However there was a big difference in the time required to complete the >>calculations at 14 ply. Apparently a Selective Search setting of 32 causes >>certain moves to be discarded more quickly with CM8666 Deep when analysis >>reaches the 14 ply level, and presumably beyond that level. >> >>=========================================== >> CM8666 Deep (Sel.Srch.=32) >> time depth score positions moves >> 10:00 13 -1.63 41072171 Rfd8... >> 28:37 14 -1.77 120083265 Rfd8... >> 1:01:20 15 -1.85 257767455 Rfd8... >> 2:54:15 16 -1.99 742496963 Rfd8... >>=========================================== >> CM8666 (Sel.Srch.=12) >> time depth score positions moves >> 9:47 13 -1.63 41072171 Rfd8... >> 27:44 14 -1.77 120082436 Rfd8... >> 2:25:54 15 -1.85 622084976 Rfd8... >> ------- 16 ----- --------- ------- >>=========================================== >> >>Most of the extra time required to reach the 15 ply level with CM8666 was spent >>in evaluating moves at the 14 ply level. After all 14 ply moves had been >>analyzed, it then took only about 22 minutes to analyze the first line at 15 ply >>as shown in the table above. It would presumably take a long time to complete >>the analysis to 16 ply with CM8666. So I aborted the CM8666 test after a total >>of 4 hours and 37 minutes. At that point it had 28 moves left to analyze out of >>43 total moves at the 15 ply level and had processed about 1,210,000,000 >>positions. >> >>................................................................... >>................................................................... >> >>Here are a couple of other observations regarding user-defined personalities: >> >>Some tests with the CM8666 and CM8666 Deep personalities with certain positions >>produced an unusual ordering of multiple evaluations within a certain ply in the >>Thinking Lines window. For example if calculating at the 10 ply level with >>black to move, the first line found and displayed may have a score of 0.50. >>Then still at 10 ply with black to move, the second line found may have a value >>of 0.75. That appears to be calculating in the wrong direction. Using the >>standard protocol which I believe most people expect, calculations with black to >>move should proceed to find smaller scores (not larger scores) at the same ply >>level! That has been explained elsewhere as normal for the CM8000 Thinking >>Lines window, and only the last move found should be considered correct, prior >>moves being "outdated". But I wasn't entirely convinced, so I tried to isolate >>the cause... >> >>I tried all of the possible CM8000 personality setting changes, one by one. >>Only changes in King Safety (a Positional setting) from the default of 100 >>caused the strange analysis data mentioned in the prior paragraph. Returning >>that setting to the default of 100 removed/prevented such strange analysis data. >> That's not conclusive and more tests are needed, but it suggests that there >>could be a problem with changing the King Safety setting from the default of >>100. >> >>In performing these various tests, it also appears that changes in user-defined >>personalities is not initialized automatically by CM8000. It is necessary to do >>something special in order to refresh the system (such as load a stored game >>containing that personality, or exit/reload CM8000, or restart the computer). >>Otherwise the old personality settings may still be used in analysis. >> >>These tests were run with a Celeron 500MHz processor, 192MB RAM, Windows 98se, >>and the typical CM8000 installation. >> >>WP > >Some comments: > >1) If you edit a personality that is currently involved in a game, your new >settings WILL NOT take effect until you start a new game with that personality. >Just because you changed the settings doesn't necessarily mean that you want to >abort the game with the old settings. > >2) As explained before, this supposed "calculating in the wrong direction" is >perfectly normal. I found a bit of text that Johan wrote that describes this >very problem (not directly, as he was actually writing to the development team >about the best way to implement the Analysis and Advice features, but he was >referring to what you are witnessing). > >White to move: > > 5 +17 57 7225 move1 ... // all seems fine > 6 -377 240 72233 move1 ... // horrible threat discovered > 6 -320 560 171444 move2 ... // stupid improvement > 6 -181 1142 371500 move3 ... // silly delay attempt > 6 -6 1766 564781 move4 ... // reasonable repair move found > >This is pretty much what you are witnessing. The eval can go "in the wrong >direction" if "move1", which APPEARED initially to be a reasonable move, turns >out to be a horribly poor move. > >I hope that clears this one up. > >3) From what I can see from Johan's documentation, any selective search values >greater than 12 are EITHER ignored (possibly meaning the default of 6 is used) >or CHANGED to 12, I'm not sure which. Therefore, I wouldn't make the assumption >that setting that value to 32 REALLY IS setting that value to 32 within The >King. > >jm JM Thanks again for your info & feedback... Changing a personality setting is an independent action, not necessarily associated with a current game. I would prefer that it initialize immediately for all purposes. I understand your info about the ordering of analysis lines, but somehow suspect additional factors are involved. Maybe, maybe not... In my hands at least, it seems to involve a change from the King Safety default setting, but that may not be (probably isn't) the whole story. My test shows that a Selective Search setting of 32 is faster than a setting of 12, after analysis reaches the 14 ply level. More tests would help to confirm it. Why? I'm not sure yet, but suspect that CM8666 Deep is acting analogous to CM6666 Deep. It's kicking in a ply later than expected, at 14 ply rather than 13 ply. I'm doubtful it would make a difference at ordinary time controls except possibly in simple endgames where deeper ply levels are reached quicker. I can't resolve my test result with your info from Johan about the 12 ply limit, which seems contrary. WP
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.