Author: Nagendra Singh Tomar
Date: 16:29:14 10/19/02
Go up one level in this thread
On October 19, 2002 at 06:29:42, Michael Neish wrote: >Help! > >I blew the dust off my program a week ago and started working on it again. I >quickly found a problem in the search routine that is causing the program to >miss some simple tactics which it really ought to see. I?d be very grateful if >someone could give me any comments. Thanks. > >I think I finally traced the problem down to the way the search routine works. >An example is probably best. > >At root I do a full-width search, with alpha = -MATE, and beta = +MATE for the >first move. This returns a value <score>. > >For subsequent moves I search using the window alpha = (score-30) and beta = >(score+30) (purely arbitrary). If the search returns a value <= (score-30) I >search again with window -MATE to (score-30), and if it returns a value >= >(score+30) I research with alpha = (score+30) to +MATE. I hope this isn?t >totally erroneous. > >Now in a certain position the full-width search will fail high at the root (good >score), so it will re-search with the window (score+30) to +MATE, but this >actually ends up finding a strong move for the opposite side and returns a fail >low second time round. Unless you are searching deeper the next time around this should not happen. the new score should always be between score+30 and MATE, because the previous serch with a smaller window failing high implied that the score is at least score+30. So I see two possibilities. 1. You are doing a deeper search the next time. 2. You are using transposition tables and this is due to that. >Since the program won?t search any more it will return >the value (score+30) thinking it?s got a good position. Then a few moves later >disaster strikes. Moreover, it somehow loses track of the PV, but I think this >may be a separate problem. > >Could someone please let me know how to deal with this sort of thing? I'm >stuck. > >Many thanks. > >Mike. regards tomar
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.