Author: martin fierz
Date: 23:38:27 09/09/02
Go up one level in this thread
On September 10, 2002 at 01:45:23, Uri Blass wrote: >On September 09, 2002 at 19:59:39, martin fierz wrote: > >>On September 09, 2002 at 18:42:28, Thorsten Czub wrote: >> >>>Only DOS :-))) >>>No tablebases for rebel. only 60 MB hash. >>>windows version hopefully released... when ed is ready :-))) >> >>x2cou_51, x2cou_56? how many versions have you tested, thorsten? >> >>i ask because i wonder if you are wasting your time with the following approach: >>if you produce 50 versions of rebel (let's imagine they are all same strong) and >>then play a short (e.g. 10 games) match against fritz with each of the 50 >>versions - what happens? thanks to statistical fluctuations *alone*, some >>versions of rebel will win, some will lose. you will think you produce a great >>new playing style, when all you do is chase random numbers... how many games do >>you play with every style? i don't want to stop you from trying to improve rebel >>- just to make sure you do it right :-) > >Note that thorsten also watch the games. >Seeing the games give more information than seeing only the results. > >Thorsten explained this point in the past. i understand - and i agree with you that if you watch the games you can learn more than just looking at results. i often go over all lost games in such a 288 game match to try and find patterns in my program's losses. i didnt know torsten was watching them closely - and so i asked what he is doing exactly. aloha martin >I believe that even with watching the games a lot of games are needed but less >games relative to the case that you do not watch the games. > >You can learn from watching the games what are the numbers that should be >changed and to what direction. > >You can learn from analyzing the games and trying to do a minimal change in the >evaluation to fix the error. > >I guess that there are cases when thorsten rejects a style inspite of having >slightly better results but only thorsten can tell if my guess is correct. > > >>let me give you an example: when i change something at my checkers engine, i >>play matches of 288 games against a control program for comparison. i have often >>seen the first 20 games be like +4 for one engine, but finally it loses the >>match... > >For me it is dependent on the change. > >There are changes in the evaluation that I do based on watching a lot of games >and analyzing them without tests and testing by games is done only after >thinking that I do a significant improvement. > >In these days I do not work on movei(I guess that I will continue in october to >work on it) but I clearly learn what should be changed from watching and >analyzing a lot of games. > >Uri > > > > > > > >> >>aloha >> martin >> >>PS: this "try lots of ideas and one will be good" strategy is something you can >>actually make money with on the stock market: an arbitrary example: use an >>M-N-strategy to make money at the stock market: at the end of every year, buy >>stock of companies which are between rank M and N in performance over the last >>year, eg M=4, N=10. the idea is to not buy companies with the best performance >>because they probably won't be able to repeat their performance, according to >>some stock market gurus. now, take 50 such strategies with different M and N >>numbers and look backward which would have performed best over the last 10 >>years. you are virtually *guaranteed* to find one which outperforms the market. >>now, the important money-making step: DO NOT USE THIS STRATEGY YOURSELF. it >>won't work, because it's just a random fluctuation thing. one of these >>strategies had to be best. instead, sell books on how to make money at the stock >>market, quoting this strategy and saying "over the last 10 years, you would have >>outperformed the market by 5%". this seems to work well :-)
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.