Author: Tony Werten
Date: 00:02:10 12/04/02
Go up one level in this thread
On December 02, 2002 at 22:33:14, Robert Hyatt wrote: >On December 02, 2002 at 18:01:00, Bob Durrett wrote: > >> >>Fritz will display second best, third best, etc. moves and give a line for each. >> But the Fritz manual says doing so requires extra computation when compared to >>calculating only one move and displaying a line for it. >> >>I do not know what other chess-playing programs have to offer insofar as display >>of alternate moves with there lines is concerned. >> >>A few questions: >> >>(1) If other programs offer this same feature, how do these chess engines >>produce the alternate moves and lines? Differently from the way the first move >>and it's line is produced? > >There are two ways to do this. > >1. Search all legal moves and get the best and display it. Delete this from >the list of legal movs and search again for the best. This will take about >as long as the first search. Total cost to find two moves is 2x the cost. > >2. search the first move and get the PV. Then as you search others, you may >get a new PV. This PV is clearly the best, and the old best is now _Probably_ >second-best. > >To do it right is expensive. > >> >>(2) If a chess-playing program were optimized to display multiple lines, would >>the search algorithms and position evaluation code be any different from those >>of a chess engine optimized for one move [and it's line] only? > >Probably not. But alpha/beta is not designed to produce anything other than >the "bst" move... > > > >> >>(3) If an engine is not optimized for this application, how much will the engine >>performance be handicapped if it is forced to display alternate moves and their >>lines? > > > >5 lines is roughly 5x slower, but it can be better or worse... For XiniX, getting the exact scores of all moves is only about 30% slower. It seems that being able to order the moves at the root on score helps the tree a lot. Plus there are a lot more exact scores from hashtable to cutoff ( I think ) I have been thinking about this since it might be worth taking this 30% hit, but have an easy multiprocessor algorithm; just splitting at the root. Ignoring the memorybandwith, this algorithm should be able to produce a nearly linear speedup ( Again, I think ) Tony > > > > > >> >>There may not be any simple answers to these questions? : ( >> >>My motive for asking these questions is obtain preliminary information which I >>might use for an idea, having to do with trees of "best move sequences," I would >>like to think more about. >> >>Bob D.
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.