Author: Mridul Muralidharan
Date: 15:16:11 03/17/04
Go up one level in this thread
On March 16, 2004 at 18:36:54, Sune Fischer wrote: >On March 16, 2004 at 17:50:02, Mridul Muralidharan wrote: > >> One additional comment which I forgot to mention : >> pcsq values based move ordering can and does perform worse than history for >>some class of positions where there are lot of "history favourable" cutoffs >>happening and for some other types of positions too. > >I would expect pcsq to be a lot worse than history sorting for most all >positions. >It takes just a second or two to get the history up and spinning, after that I >don't see how pcsq is going to compete. > >> But in most of these cases , using ply - 2 killers and bumping number of >>killers from 2 to 3 or say 4 was also sufficient compensation for removing >>history even these sort of positions ! >> >> Take a set of quiet positions and it will be possible to see history being >>totally equal or slightly worse or slightly better than random (or no) move >>ordering. > >I don't think so, random is about as bad as you can get. > >History tables, while they may not be perfect, are certainly an improvement upon >randomness. Maybe I forgot to add YMMV :) But my expiriments on crafty have lead me to believe that crafty can do better than history. Ofcourse , for me , history totally sucks - no doubt about it. History is damn cheap (well almost :) ) , and almost all sites on chess programming have it mentioned ;) so anyone starting off would think it is like a gospel truths that history performs well. The amount of attention that it gets and the amount of chess programs that it has made its way into is dispropotionate to the benifit it gives :) If people expiriment more (and more often too !) , then you may come up with other schemes that are far better. I dont use history or pcsq or any crude piece/from-to based move ordering schemes anymore - gave them up about 6 months or so after starting seriously. I still keep revisiting them , tweaking them , etc due the obvious "cheapness" in these schemes - IF i can get a good solution - then why not use it ! ;) But of all the initial attempts , I found history totally bad and nothing has changed my opinion about this till date. (Even the very interesting study by someone a few months (?) back on various move ordering heuretics was not really convincing since I did the same and found contrary results .... maybe mine of his code had bugs ;) ) Maybe in initial years of computer chess when processing power was very very pricy then history might have had some kind of benifit (whenever with luck it does not deteriorate into randomness ...) but harping about same ideas 10 - 25 years later still does not smack well of progressive and scientific thinking .... Today , you can definitely attempt better schemes - a 50 knps drop if it leads to 20 % smaller tree is an amazing gain today - while such a propotionate speed drop when programs were barely doing 6 - 8 ply was unthinkable earlier on ! (reasons for this should be obvious ...) Ok , enough rambling , 3:30 at night can do crazy things to brain :( Good night , Best Regards and have a "happy history" ;) Mridul > >-S. >> Eliminating all these and making move ordering as good as possible made >>messchess a very slow engine that it is now. (This and a lot of other search and >>eval related stuff :) ). >> >>Mridul
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.