Author: Matthew Hull
Date: 06:34:14 03/16/04
Go up one level in this thread
On March 16, 2004 at 06:32:55, Steven Edwards wrote: >On March 16, 2004 at 04:51:32, Vasik Rajlich wrote: >>On March 16, 2004 at 03:08:06, Dave Gomboc wrote: >>>On March 15, 2004 at 17:23:32, Steven Edwards wrote: >>>>On March 15, 2004 at 16:52:40, martin fierz wrote: >>>>>On March 15, 2004 at 16:38:53, Steven Edwards wrote: >>>> >>>>>>Hmmn. Maybe I should offer a wager or two here to the doubters. Like, if I >>>>>>can't get this to work, then I'll stop complaining about the mundane nature of >>>>>>traditional A/B searchers; if I do get it too work, each doubter can send me a >>>>>>new battery for one of my Macintosh notebooks. (Approx. US$150 each.) >>>>> >>>>>i'll accept the wager, but you have to define "can't get this to work" a bit >>>>>more clearly for me. e.g. IIRC your list had an item "become world champion", >>>>>and i would accept that you had "got it to work" long before that. >>>>>for me, the getting it to work part has to be spelled out as some kind of rating >>>>>level - what do you think? what level would you specify? >>>> >>>>Well, first let's hope our board sponsor won't get upset with a little gambling. >>>> >>>>I posted the primary and secondary goals back last month but can't find the CCC >>>>reference. So you are welcome to read them again from the entry 2004.02.19 in >>>>my journal: >>>> >>>>http://www.livejournal.com/users/chessnotation/ >>>> >>>>Point #8 in the primary goal set (combined with #12) is what I claim to be >>>>sufficient for proof of concept, and I'll make the output public for inspection. >>> >>>Unlike some of the other posters, I'm less skeptical as to what you'll be able >>>to come up with using this approach if you follow it through fully, but I do >>>think that your targeted playing strength (1800 elo) is unreasonably low. It's >>>acceptable for your program to be weaker than top programs, which are on today's >>>hardware approaching 2700 strength, but not by such a wide difference. And in >>>it's day, Paradise was able to handle its domain at expert strength. >>> >>>I suggest that your rating target be 2400 elo. Really. You can do it. With >>>regard to a possible bet between you and Martin Fierz, 2200 on today's hardware >>>seems like a fair level. > >>I'd be ready to contribute to the "bet", but everything would have to be >>concrete. >> >>#8 is vague - if I wanted, Rybka could be meeting this tonight. #12 - what test >>suite are we talking about? WAC? In addition, you won't be helping your engine >>if the target is testsuite performance - good for testsuites != good for >>positional play. >> >>2400 elo? Shredder & Hiarcs search 300 KNPS @ 3 GHz and play ~2800. Slow them >>down by a factor of 300, log (2) (300) = ~8, 8x50 ELO/speed doubling = 400 >>points - so ~2400 @ 1000 NPS. (Actually 50 is probably too low, but still.) >> >>I can join if the goal is 2600 ELO @<1000 NPS. That's proof of concept. In this >>case, I think you could find a few contributors ... > >On the test suite data: > >I'll claim that #19 is satisfied if Symbolic can solve at least 200 of WAC, 667 >of WCSAC, and 667 of BWTC with a mean time limit of 180 seconds per position on >hardware roughly equal to #11 (400 MHz PPC with 256 MByte RAM and 10 GByte >disk). > >A few points on the 1800 Elo number in the primary goal set: > >1. The only test suite data we have for Paradise is just under a hundred >positions from the first one hundred positions from WAC. I think its >performance was somewhat under 2000 Elo because of its time limit of forty-five >minutes per move, and that is one reason for the 1800 Elo figure in the primary >goal set. > >2. A second reason for the 1800 Elo figure is that I suspect that, unlike the >case with most programs, incremental improvement is going to be strongly >correlated with the chess skill of the improver (me). I haven't played OTB >chess in some 20 years and I don't think I could do much better than 1800 if I >were to try it today. > >3. The third reason for the 1800 number is how I would like to get the rating: >from OTB USCF tournaments. The human player distribution is such that there >would be a lot more confidence in the target number when the number is within >two sigma of the mean and not four (or more) away. > >4. A difficulty here is getting any kind of an OTB rating. Can you not get the "OTB" equivalent on FICS or ICC? You can limit opponents to humans. There are many 1800 and below sparring programs that run on these servers that cater to humans. >To my knowledge, >there haven't been any "computers allowed" USCF events in my area for a long >time. There were a good number back in the late 1980s when I deployed my >program Spector, but not today. My idea here is to get some local TDs to allow >computer entry by helping to sponsor a prize fund, and this might be a budget >breaker. (Note: this is the reason for the caveat in #18 and #19.) > >-------- > >On the battery: one of my notebooks is a 400 MHz Apple PowerBook G3 2000 Series >and a battery for it can be had from: > >http://eshop.macsales.com/Catalog_Item.cfm?ID=5833 > >I need a couple of other batteries, all roughly the same price.
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.