Author: Bruce Moreland
Date: 11:08:23 04/10/98
Go up one level in this thread
On April 10, 1998 at 13:25:01, Moritz Berger wrote: >I don't think that it is that easy. Just compare the "mean and lean" 54 >KB chess engine of the King 2.54 with Fritz' 105 KB "twice as much >knowledge" engine. Both programs play good chess. I know that some of >you will not agree with this quantitative approach (the Hiarcs 6 engine >is BTW about 200 KB big), but chess knowledge (whatever that might be) >has to be coded somewhere and counting engine sizes gives at least some >metrics to compare "amount of knowledge". Nodes per second (NPS) is >another quantum, useful when trying to compare "expensive" (i.e. higher >order) and "inexpensive" (e.g. root node only evaluated, preprocessed) >knowledge. You can't compare EXE sizes. I can make my exe way bigger by including a lot of pre-initialized data. I can make it smaller by taking this out and generating the data at boot time. I can make the EXE bigger by providing several routines that do almost exactly the same thing. I can make it smaller by having one routine that does conditional branches all over the place. All without modifying the evaluation function at all. >The essence of this argument IMHO is that there can't be *THAT* much a >difference in chess knowledge between engines that are of the same >magnitude in size and NPS (e.g. Fritz does only about 2-2.5 times as >many NPS as Rebel, even if NPS are not always comparable, but they are >within the same oder of magnitude). If you got a 20% raise, does this mean you could eat in restaurants 20% more often? No, you could eat in restaurants every day, assuming you didn't expand every other aspect of your spending by 20%. If you are willing to drop your NPS by 20%, you can do a heck of a lot more eval, assuming that all of the drop comes from increased eval. A factor of 2 allows you to do a lot more. But honestly I don't know exactly what people are using the time for. Fritz seems to be not only fast, but efficient. I don't know what percentage of the time other programs are spending trying to do search extensions (this isn't eval function knowledge), as opposed to how much they are spending trying to understand how good a minor piece is in its specific context (this is eval function knowledge), or what. I don't know what the deal is with Fritz, Frans is kind of closed so it is kind of hard to know what he is really doing, so it is very hard to even wonder what is going on, and much harder to make real *conclusions*. The important thing is to look at what it achieves, and forget the NPS number. I don't care whether he is spitting out 1,000,000 or 1,000, NPS is an easy number to see but people often look at it and stop thinking. What matters is how the thing plays, and there is no solid-gold 100% accurate number associated with this. I haven't tried to watch Fritz play, but when people use Fritz 5 on ICC against my program, I see more "-" in my result column than I did before. >Big words ... In the current CSS issue you will read the Kasparov >favours the Fritz 5 engine over Hiarcs for analysis (which he had >previously recommended as being better than the rest of the pack). I don't have anything against CSS, but on the other hand I wouldn't listen very seriously to Michael Jordan telling people about how great Nike is. 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.