Author: Komputer Korner
Date: 09:13:50 06/30/98
Go up one level in this thread
On June 29, 1998 at 14:17:55, blass uri wrote: >On June 29, 1998 at 13:55:30, Komputer Korner wrote: > >>On June 29, 1998 at 11:07:13, blass uri wrote: >> >>>Fritz5 has a function of correspondence analysis. >>>with this function Fritz5 generates a detailed analysis tree for a position. >>>I can decide how deep and broad the resulting tree should get and what >>>moves should be included or excluded from the analysis. >>> >>>in order to use the fritz5 correspondence analysis I give fritz5 some numbers: >>>1)time per move(it is time per node in the tree and it is not a good idea >>>because I want to give the computer more time to analyze the nodes near the root >>>because it is more important not to do a mistake in analysis near the root I >>>think the possibility to choose time per ply 1,time per ply 2,... is a better >>>idea) >>> >>>2)Branching:The branching factors for the 1st,2nd and 3rd moves define >>>how many alternative moves are given in the first ply,second ply and third ply >>> >>>3)Length of variation:This sets the length of analysed variations >>> >>>4)Evaluation window:This value ,given in 100ths of pawns,determine when >>>Fritz will discard a variation. If you set the value to 80, for instance, >>>then moves which are 0.8 pawns worse than the best move will not be investigated >>>any further-even if the branching factor would allow it. >>> >>>The main problem with the tree fritz5 generates is the fact fritz5 >>>analyses illogical lines >>> >>>for example suppose the branching factors are 2 for the 1st ply 2 for the >>>second ply and 1 for the third ply >>>and the length of variation is 3 >>>and the time number is 60 seconds per move >>> >>>Fritz5 can do the following tree >>>it is only an example(fritz5 did not generate this tree but generates trees >>>with the same mistake) >>> >>>line 1:1.e4 e5 Nf3 0.1 pawns for white >>>line 2:1.e4 c5 Nf3 0.2 pawns for white >>>line 3:1.d4 Nf6 c4 0.08 pawns for white >>>line 4:1.d4 d5 2.Nf3 0.04 pawns for white. >>> >>>near 1 minute for finding 1.e4 and 1.d4 the two best moves. >>>near 1 minute for finding 1...e5 and 1...c5(2 best replies to 1.e4) >>>near 1 minute for finding 1...Nf6 and 1...d5(2 best replies to 1.d4) >>>near 4 minute for finding the third ply in lines 1,2,3,4(1 minute for a line) >>> >>>I do not need the last line to understand 1.e4 is better than 1.d4 >>>because I know the value of 1.e4 is 0.1 pawns for white >>>and I know the value of 1.d4 is at most 0.08 pawns for white but fritz5 >>>analyse the position after 1...d5 and waste time about it. >>> >>>fritz5 does not use the alpha beta for the tree it shows me and I do not >>>understand why. >> >>When you are in k variation mode which is essentially what the correspondence >>feature is, alpha beta only applies within each variation, not across >>variations. > >I think it should be simple to use the alpha beta across variations. >fritz5 uses the alpha beta for a very large tree I cannot see >and it should be simple to use the alpha beta for the tree I see >when the computer computes near 1 node per minute in this tree >instead of 200000 nodes per second. > >Uri Just how do you propose to do this? Alpha beta doesn't find an exact score except for the best move. Forcing the program to find an exact score for every candidate move slows the program down immensely. Using mini maxing across variations is only possible if the variations are part of one big search and then you are back to having an exact score for each candidate move which is too slow in practice. -- Komputer Korner
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.