Author: martin fierz
Date: 13:23:21 03/30/02
Go up one level in this thread
On March 29, 2002 at 19:39:43, Christophe Theron wrote: >On March 29, 2002 at 07:33:34, Ed Schröder wrote: > >>Dear CCC, >> >>This posting is dedicated to Thorsten in the first place being the absolute >>opponent against statistics. He might change opinions after this conclusive >>posting, Thorsten you will love this one! >> >>It's also dedicated to those who still believe that there is no performance >>increase/decrease when playing on different time controls. (Hi Christophe!) > > > >That's not exactly what I believe. > >I believe that it is possible to write a chess program that will be the best at >any time control. > >That's the direction I am try to follow, and that's why I take into account the >performance of my program at all time controls, including blitz on very slow >hardware (old 386 or Palms). > >I also consider that a program that is good only at a given time control has a >serious problem. > > > > Christophe here's an example: let's say, you write a new chess program and implement some move ordering. it runs at some speed, X kN/s. now, you improve your move ordering, with a very elaborate scheme which makes it run at only X/2 kN/s. but your move ordering is better, which means the exponent a in the search tree size (nodes(d) ~ k^(a*d)) is smaller than before. if your new a is only a bit smaller than the old one, you will need a deep search to justify your new move ordering scheme. if this example is correct, and assuming that you can choose your move ordering to be either good and slow or fast and bad, and can vary it between these extremes, then there IS an optimal setting for your program for a given search time. i think this also shows that ed schröder's original post is not necessarily correct - the program with the best move ordering should gain most from increasing hardware speed. now according to his post, rebel gains the most - but this does not mean that it will be best on the 3600MHz computers, because the other guys can increase their efforts in move ordering as computers get faster. i agree that this argument is misused for weak programs. i'm not trying to defend them :-) i'm just trying to justify that my checkers move generator needs about 30% of my program's execution time - only for move ordering. i usually run a test suite at depth 13 and depth 19 to test changes in move ordering, and if you do something better, you often see that it needs e.g. 5% less nodes on average to get to depth 13 but 10% less on average to get to depth 19. now if it is 7% slower with the additional ordering it's not a good idea if the typical search depth is 13, but a good idea if it's 19. it's some time ago since i last revisited my move ordering, but the numbers are realistic. aloha martin > > > >>The below data is taken from the latest SSDF list and measures the differences >>between the used hardware, that is Athlon 1200 Mhz versus the AMD 450, roughly >>being a factor of 3. >> >> 2 Chess Tiger 14.0 CB 256MB Athlon 1200 2714 34 -33 460 63% >>2619 +82 >> 13 Chess Tiger 14.0 CB 128MB K6-2 450 MHz 2632 30 -29 585 65% 2526 >> >> 3 Deep Fritz 256MB Athlon 1200 MHz 2709 34 -33 447 63% >>2617 +55 >> 9 Deep Fritz 128MB K6-2 450 MHz 2654 24 -23 901 62% 2568 >> >> 4 Gambit Tiger 2.0 256MB Athlon 1200 2708 37 -36 381 61% >>2626 +67 >> 11 Gambit Tiger 2.0 128MB K6-2 450 MHz 2641 30 -29 592 66% 2522 >> >> 6 Junior 7.0 256MB Athlon 1200 MHz 2684 34 -33 436 58% >>2630 +53 >> 14 Junior 7.0 128MB K6-2 450 MHz 2631 28 -26 717 67% 2505 >> >> 7 Rebel Century 4.0 256MB Athlon 1200 MHz 2680 36 -35 407 63% >>2587 +104 >> 21 Rebel Century 4.0 128MB K6-2 450 MHz 2576 68 -70 105 46% 2606 >> >> 8 Shredder 5.32 256MB Athlon 1200 MHz 2662 32 -31 498 56% >>2619 +56 >> 17 Shredder 5.32 128MB K6-2 450 MHz 2606 31 -30 545 62% 2521 >> >> 10 Gandalf 4.32h 256MB Athlon 1200 MHz 2649 34 -33 430 54% >>2621 +130 >> 32 Gandalf 4.32h 128MB K6-2 450 MHz 2519 36 -36 378 53% 2499 >> >>Based on the above comparison the 2004 SSDF list running on 3600 Mhz hardware >>might looks as follows: >> >> 1. Chess Tiger 14.0 CB 256MB Athlon 1200 2714 + 82 = 2796 >> 2. Rebel Century 4.0 256MB Athlon 1200 MHz 2680 + 104 = 2784 >> 3. Gandalf 4.32h 256MB Athlon 1200 MHz 2649 + 130 = 2779 >> 4. Gambit Tiger 2.0 256MB Athlon 1200 2708 + 67 = 2777 >> 5. Deep Fritz 256MB Athlon 1200 MHz 2709 + 55 = 2764 >> 6. Junior 7.0 256MB Athlon 1200 MHz 2684 + 53 = 2737 >> 7. Shredder 5.32 256MB Athlon 1200 MHz 2662 + 56 = 2718 >> >>I love this list :) >> >>Unknown entries because of no comparison material. >> >> 1 Fritz 7.0 256MB Athlon 1200 MHz 2748 38 -35 395 68% 2616 >> 5 Shredder 6.0 256MB Athlon 1200 MHz 2706 37 -36 379 61% 2626 >> 16 Crafty 18.12/CB 256MB Athlon 1200 MHz 2613 33 -32 459 54% 2587 >> 23 Nimzo 8.0 128MB K6-2 450 MHz 2554 24 -24 824 53% 2531 >> 29 Hiarcs 7.32 128MB K6-2 450 MHz 2526 21 -22 1051 48% 2537 >> 34 Chessmaster 8000 128MB K6-2 450 MHz 2517 44 -45 251 45% 2550 >> >>A couple of notes: >> >>1) Rebel Century 4 (AMD-450) has played to less games yet (105) to make it a >>reliable comparison. However the result (+104) doesn't surprise me, Rebel is >>tuned for longer time controls. >> >>2) It has to be seen if Fritz7 and Shredder6 have not improved playing on faster >>hardware. Since there is no comparison material yet this remains unknown for the >>moment. >> >>3) I don't think the above is a scientific conclusion, neither a 100% valid >>prediction, but IMHO it isn't 100% nonsense also. >> >>4) You are invited to shoot the above into pieces. >> >>Ed
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.