Author: Ed Schröder
Date: 06:52:35 12/30/97
Go up one level in this thread
>Posted by Thorsten Czub on December 29, 1997 at 18:04:54:
>Please, my analysis have been done on a k6/200 with giving Genius5 3968
>KByte for hash.
>For point 1 use pgn1, for point2 use pgn2:
>[Event "PGN1"]
>[Site "?"]
>[Date "????.??.??"]
>[Round "?"]
>[White "Karpov"]
>[Black "Topalov"]
>[Result "1-0"]
>[SetUp "1"]
>[FEN "rqr3k1/3Qbp2/p1n1p1p1/1pp5/2P2P2/2N3P1/PP3PB1/R3R1K1 w - -"]
>1. *
>[Event "PGN2"]
>[Site "?"]
>[Date "????.??.??"]
>[Round "?"]
>[White "Karpov"]
>[Black "Topalov"]
>[Result "*"]
>[SetUp "1"]
>[FEN "rq3rk1/3Qbp2/p1n1p1p1/1pp5/2P2P2/2N3P1/PP3PB1/R3R1K1 b - -"]
>1. *
>1.
>Set up the position PGN1 in Genius5/WIN (import it as PGN or setup
>manually).
>Put level on infinite.
>Let Genius5 compute on the whites move.
>It considers Bxc6 in the 8/20 search.
>Changes to cxb5 in this search.
>It runs through the 9/21 search without changes until (arround 51' !!)
>it computes in the 34th
>branch (that is Rxe6), but did not see the point and starts the next
>iteration (10/22) with main-line
>(since 18'58") cxb5 Ra7 Qd3 Nd4 bxa6 Qxb2 Rab1 +0.60.
It's not a good example IMO.
When I set the G5 playing style to "Risky" Rxe6 is quickly found.
I think the G5 score differences between Bxc6 and Rxe6 are to
narrow to proof your point with this position.
Besides of that every selective program (null-move) or otherwise
has errors in pruning the tree. There will be no algorithm ever
which will replace the brute force approach. Just set any Rebel
version to brute force and watch the differences. Very shocking
sometimes...
The same applies for every program I have seen which offers the
user the "brute force" option. It shows you the weak points in the
selective search of a program.
I am collecting these kind of failures since many years. In every
new Rebel version I try to fix one or two patterns. It's a
nightmare for every chess programmer.
- Ed -
>2.
>Now set up again the position but take the rook from c to f8 !
>Or import PGN2, that is the old position but with the rook-move taken
>back.
>Let Genius5 compute about THIS position that is ONE ply before the
>unique key-move Rxe6 with
>following Rxg6.
>After 1'23 in only search 8/20 Genius5 finds the key move Rxe6 , but
>here of course as the 2nd
>ply of the main-line !
>The main line is:
>Rc8 Rxe6 Ra7 Rxg6+ fxg6 Qxe6+ Kg7 Bxc6 with an evaluation of -1.15.
>It will still play Rc8 in the 9/21 search with -1.15 evaluation.
>Conclusions:
>This is what I call the asymmetric-search. It is the reason Genius5 is
>that strong and plays that
>"Levy"-like ("do nothing, but do it well").
>The reason it finds the key moves FASTER although you go back (!!) one
>ply shows that it
>computes more brute-force (less pruning) in the plies 2,4,6,8,... !
>The reason it has not found Rxe6 in the position with the rook on c8 is,
>that is had to find the key
>move with the plies that prune much and have pruned it away !!
>ALL Lang program have this search-technique.
>>From the first dedicated machine Mephisto Amsterdam to the latest
>Genius5 engine.
>Lang has changed the knowledge and the tactical abilities, but never has
>he changed the search-
>strategy.
>
>I will further (when I have some time) show you a position where Genius5
>decides to choose the
>boring (also weaker) moves/main-lines with having 1,3,5,7...
>plies-search, and if you take one
>ply back, it plays for the same color better, more active moves that
>would lead to a more
>powerful game. It IS the search that makes genius play "boring" since
>1985 Amsterdam-module.
>Thanks.
>P.S.: Take care about those effects when analysing your own games with
>Genius5. Especially
>your email-chess-games ! You cannot trust Genius' analysis even if you
>let it compute 12 hours
>per move. The effect seen in the position above will not disappear
>whatever time you give it. All
>your analysis time invested could be WASTED because the FIRST move in
>the main-line is NOT
>the best one in the position you analyse.
>Never trust genius5's 1,3,5,7... iteration-moves from the main-line !
>They could be completely
>overseeing STRONG moves.
>Use Genius as an instrument. But use it not blind trusting. Remember the
>asymmetric search
>and take care of it.
>I remember a very old "argument" with Thomas Mally or Andreas Mader
>(both writers) from PC-
>SCHACH/MODUL in the past. He said that he cannot understand my sceptical
>position
>concerning using the dedicated mephisto-machines (with Lang program) for
>analysis. In fact it
>was the Mephisto Lyon or Almeria or something like this.
>He told me that he even bought a second machine for analysis !!
>I was very shocked. I still remember this episode, although it is years
>ago.
>Sitting with Genius5 about the above position, having much faster
>hardware than the old
>Motorola 68000 or 68020 CPU, I am still convinced that my decision NOT
>to trust those machines
>due to their lack of "accuracy" that results from the highly contrast
>between the 2 pruning-levels
>was reasonable.
>If a program is not able to find such an important move although it runs
>on such a fast machine,
>how much would it would have overseen on a (how many times?) slower
>Motorola 68020 ??!!
This page took 0.01 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.