Author: Anthony Cozzie
Date: 07:27:19 03/29/04
Go up one level in this thread
On March 29, 2004 at 10:01:55, martin fierz wrote: >On March 29, 2004 at 09:46:48, Uri Blass wrote: > >>On March 29, 2004 at 08:54:09, martin fierz wrote: >> >>>On March 29, 2004 at 07:04:13, Uri Blass wrote: >>> >>>>On March 29, 2004 at 06:25:28, martin fierz wrote: >>>> >>>>>On March 28, 2004 at 18:55:53, Artem Pyatakov wrote: >>>>> >>>>>hi artem, >>>>> >>>>>[snip] >>>>> >>>>>>>>On the other hand, I think a lot of researchers have been overly ambitious and >>>>>>>>have tried to replace Alpha-Beta & tricks with a neural network or some totally >>>>>>>>different approach. I think that with the current state of AI tools, these >>>>>>>>efforts are bound to fail. >>>>>>> >>>>>>>A lot of researchers? Other than myself, I don't know of any other workers who >>>>>>>are attempting a complete and competitive chess playing program that doesn't >>>>>>>tread the oft traveled A/B bag-of-tricks road. >>>>>> >>>>>>That's good a point (will probably include in paper) - not many are working on a >>>>>>complete chess program, since most just address a particular section, like the >>>>>>evaluation function. But one old effort - MORPH - comes to mind, and I think I >>>>>>noticed a couple of other failed Neural Net efforts (don't have references >>>>>>handy, but I can look). >>>>> >>>>>steven is doing a completely radical departure from A/B. but many others are >>>>>applying more gradual changes to the A/B-model, with reasonable success >>>>>(smarthink, kaissa, gothmog come to mind, and probably most commercial engines >>>>>too). at least these engines make considerably larger attempts to shape their >>>>>game tree than a "classic" A/B searcher like crafty does. >>>>> >>>>>[snip] >>>>> >>>>>>As I mentioned before, I decided to concentrate my research on move ordering for >>>>>>a couple of reasons: >>>>>>1) Move ordering promises big payoffs if done right. >>>>> >>>>>how much deeper can you search if you reach 95% move ordering success (defined >>>>>as cutoff @ 1st / all cutoffs) instead of ~90% that most get today? how much if >>>>>you reach 100%? >>>> >>>>I cannot answer the 95% question but I can answer the 100%. >>>> >>>>If you reach 100% then always choosing the first move is enough because if the >>>>first move does not generate a cutoff no move is going to generate a cutoff so >>>>you can finish to search to the end of the game very quickly because you search >>>>only one move in every ply. >>>> >>>>Uri >>> >>>you do not answer the question here - and what you answer is wrong too :-) >>>always choosing the best move first to search means you search sqrt(N) moves per >>>ply on average, not 1 move per ply. >> >>No >> >>In that case you do not need to search the other moves because you know that >>they are worse. >> >>If you always start from the best move than the pruning strategy of not >>searching moves except the first move is simply correct and I see no reason not >>to use it. >> >>Uri > >come on uri, this is ridiculous. of course you can use that shortcut IF you know >that your move ordering is 100% perfect. in that case you need no search at all >either. obviously this is an academic question and has nothing to do with >reality, and is not what i'm asking... the question remains: in the extremely >unlikely case that you have 100% perfect move ordering (by accident), what do >you gain compared to 90% perfect (in the usual sense of move ordering %age)? > >cheers > martin It would be possible to determine this experimentally - just keep track of each position and the best move at each ply on the hard disk. It wouldn't be practical for use but it would be pretty interesting. anthony
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.