Author: Rolf Tueschen
Date: 03:22:39 01/21/03
Go up one level in this thread
On January 20, 2003 at 19:04:31, James T. Walker wrote: >On January 20, 2003 at 18:10:49, Robert Hyatt wrote: > >>On January 20, 2003 at 16:20:35, James T. Walker wrote: >> >>>On January 20, 2003 at 15:25:20, Robert Hyatt wrote: >>> >>>>On January 20, 2003 at 11:39:27, James T. Walker wrote: >>>> >>>>>Neither will 90 rounds. I've seen some discussion about the >>>>>times/rounds/playoffs of CCT mostly looking for ways to improve the format. In >>>>>my opinion as a spectator the format is great. I even liked the playoff format. >>>>> I believe a world championship was decided in a similiar manner not too long >>>>>ago. Nobody should expect a swiss system event to produce the strongest player >>>>>as the winner every time. However in my opinion this was the case this time. >>>>>I'm also curious about some programmers claiming the blitz playoff is not good >>>>>because their program is tuned for longer time controls. I wonder how you do >>>>>that. I mean if you are playing your program on ICC for games, how does playing >>>>>80% or more of your games at blitz/lightning help you to tune for 40/2? Why >>>>>would you want your program to perform better at 40/2 than at G/5 compared to >>>>>other engines? It seems to me that the SSDF is one of the few organizations >>>>>still using 40/2 for comparison. I see this as an outdated idea. The trend is >>>>>toward faster time controls to better serve the spectators interest. All this >>>>>is from a non programming spectator so don't give it much thought. >>>>>:-) >>>>>Jim >>>> >>>> >>>>You miss the point. A _tournament_ will _never_ give you the "best program." >>>> >>>>It will just give you a winner, hopefully. There is a big difference between >>>>"winner" and "best program". The difference can be explained statistically, if >>>>you are interested... >>> >>>How did I miss the point since that was the topic of my post? I suspect you >>>missed the point. >>>Jim >> >> >>OK.. perhaps I misunderstood your first "implication". "neither will 90 games." >> >>That seems to be leading into the old "you need a _bunch_ of games to separate >>two programs that are very close" discussion. And no tournament ever claims to >>identify the "best player". They just identify "the winner" > >Yes. Exactly. And so all I'm saying is that there is no tournament system which >will prove who is best only who won this time. The Swiss system is good for >what you guys were doing and I like the format as a spectator. So I'm really >campaigning to just keep it the way it is and let all programs hope for a better >result next time. >Crafty played very well by the way and I really enjoyed the Crafty Ruffian game. >Jim Let me try to find the reason for the apparent oversight of our expert number 1. I think I know the reason, Jim, excuse me. Because you are simply not right with your "neither in 90 games". That is statistics. Now Bob says that you can't seperate two progs that are closely together even NOT with a bunch of games, because that would be the typical lay error in stats. Which is completely wrong as I already had explained to Dann Corbit. I hope that I can explain that, because it's important. If you have two progs and they closely together with their Elo number then that means EITHER that they are really that closely together, or they are not even different in Elo numbers that means zero difference or they are 100 points apart, just to give a number - it is also to be seen in the given range in the SSDF results for example, and that would only come out after a bunch of games. We have these three possibilities! It's Bob who made a classical mistake, because he's so used to this reasoning that with a few games more you can't win more certainty, which is correct, that he extrapolates that "connection of uncertainty" into the infinite, which is wrong. In real it is clear, at least for chess, we are not tossing coins (Corbit), that we would have a clear seperation if the progs had any differences. Ok, if NOT, then the result would go closer to zero more and more. Of course, all this has NOTHING to do with the term "better". Because better can't remain a statistical notion, but it must be defined in terms of chess, and that would simply be silly to say that a prog with provenly 10 points more would play better chess. In the case of computerchess, as a lay myself, I would even say that a difference of 80 would not mean that chess related such a program is better. But I would really like to see such a result caused by some changes in the code if some programmer could show that e.g. by neutralizing his - say free pawn feature - a prog would become 30 points weaker or such some. Anyone on that one? Thanks in advance. Regards, Rolf Tueschen
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.