Author: Bruce Moreland
Date: 10:30:22 12/30/98
Go up one level in this thread
On December 30, 1998 at 08:33:08, Frank Schneider wrote: >Some more results: > >correct / nodes |n=infinite |n=15 |n=10 |n=5 >-------------------------------------------------------------------------------- >Standard #6 |23 / 8306851 |22 / 6616146 |25 / 6229700 |24 / 5852685 >WAC #6 |256 / 13428654 |254 / 11575072 |253 / 10977570 |253 / 10987063 >-------------------------------------------------------------------------------- >Standard #7 |25 / 26399528 |20 / 20759570 | | >WAC #7 |280 / 35352959 |274 / 30579448 | | >-------------------------------------------------------------------------------- >As expected it doesn't look good anymore. Are there possible refinements? I don't think you are looking at the right thing, so I don't think you can tell, yet. Obviously what you want to be doing is seeing if there is an improvement in real games, but this is very hard to measure, I think. I think that if you want to test tactical speed, which is what I think you want here, what you should is run this against a big tactical suite, for a fixed *time* per position. What you are doing is running for a fixed depth, so what you've shown is that you don't get as many right if you give your new version like 20-25% less time than the old version. A fixed-depth comparison will tend to discriminate against pruning heuristics and reward extension heuristics. Also, I think it makes sense to break things down into one-second bins. Here is an example, a comparison between my program from about a year ago, and my most recent version, using the ECM suite run for 20 seconds per position: #### 378 524 ---- ---- ---- 0001 375 408 0002 428 465 0003 458 497 0004 481 516 0005 503 539 0006 509 549 0007 519 564 0008 533 571 0009 540 580 0010 549 583 0011 553 589 0012 558 593 0013 563 596 0014 571 601 0015 575 606 0016 581 612 0017 587 614 0018 589 619 0019 594 623 0020 600 625 It's very clear from this that the newer one is tactically faster, much more clear, in my opinion, than just comparing 600 with 625. bruce
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.