Author: Dann Corbit
Date: 21:14:57 07/15/00
Go up one level in this thread
On July 15, 2000 at 23:39:57, Stephen A. Boak wrote: >>Simplifying. I have a penny. >>I toss it twice. >>Heads, heads. >>I toss it twice >>Heads, heads. >>I toss it twice >>Tails, heads. >>I toss it twice >>Heads, tails. >> >>I count them up. >> >>Heads are stronger than tails. >> >>My conclusion is faulty. Why? Because I did not gather enough data. >> >>In the case of chess games, as the GM's learn from previous mistakes, the true >>strength of the program will be revealed. The same would be true of a human >>player. Given enough games, they will settle down to their true strength. >> >>With a computer, which is more deterministic than a human and if it learns, does >>not learn nearly so well, the effect of more games will be far more devastating. >> >>At least according to the model I am imagining. But models can be wrong, or >>imprecise. > >Ho hum. > >Here's some food for thought. Not all may be highly digestable! > >1. If Junior6 plays 9 games against top grandmasters under tournament >conditions, the results are (some) evidence of its strength. Period. Certainly >the results are more evidence than existed before Junior6 played those >grandmasters. absolutely no one disputes this. >The weight that should be given to those results is subject to debate and >interpretation and will likely lead to differences of opinion, even opinion >well-reasoned on different sides. But evidence it is. > >How much evidence is needed to draw a conclusion? a mathematical conclusion? >When will a critic be convinced that a program is GM strength? Is there a >mathematical confidence level that *all* critics will accept? Is that enough? >See next point. when a computer accomplishes all the demands for a normal gm (fide sanctioned or not) I will admit that is good enough for me. >2. If enough trials are required to determine if, for example, a tossed coin is >fair, how many trials is that? If 100 tosses is not enough, how many tosses >would be enough? This is debatable also. It's a function of confidence level. Some may not be convinced at a confidence level of 99.88%, but 2/3 would be good enough for me. >If statistical uncertainty exists, no matter how many trials are conducted, does >that mean certainty is never possible according to mathematics? I wouldn't >continually pound sand to say there are not enough trials to draw a conclusion, >if the math I relied on would never allow a certain conclusion to be drawn, >strictly speaking in the language of mathematics. After all, with a fair coin >the possibility of 1,000,000 tosses resulting in all heads is possible--even >with the fair coin, right? Right now, a single standard deviation renders the data impotent. >3. As to the belief that a GM would (likely) learn from a series of computer >games, and the program would (likely) not learn the equivalent (about the GM's >weaknesses): > > A. If, say, Kramnik studied carefully all the games of Akopian, and >thereafter never lost a single game to Akopian but rather won 80% of the points, >would that make Akopian *not* a GM? No. Agree, but if any other GM could use the same technique and beat him, he would no longer be of GM strength. >Would that mean that *all* GMs could >accomplish the same feat as Kramnik? No. > > B. If modern GMs, having the advantage of many, many games of predecessors to >study, would be able to study the games of a very strong player of the past, >wouldn't it be likely that the modern GMs would be able to beat those historical >figures--if a match could be arranged today between the same individuals--the >modern figure and the historical one? After all, they have been able to study >the historical player's games, but not vice versa. Angels on heads of pins make for interesting debate. If you can teleport people back and forth in time, I can train them with modern methods. Then we'll see. >Would this make the older GMs not GMs after all? > >I don't think this is an automatic 'Yes' answer. The older players might have >been able to learn and beat the modern players because they are also humans, and >might adapt to modern play very quickly (learning also after some games). > > C. If a computer program is a 'learning' program, but the learning curve was >slower (initially) than the GM opponents learning curves, would the fact that >the humans quickly jumped on the computers weaknesses (choice of openings, etc) >mean that the humans would *always* dominate the computers thereafter, or might >the learning curve of the computer eventually surpass that of the humans, under >some circumstances? Iteresting question. I don't think we have the data to answer it. Do you? > D. If a GM ages and then fails to achieve the same results as when he was >younger and played better (relative to his opponents, of course), was he not >nevertheless a GM in his younger days, although his opponents continue to learn >and eventually best him a lot more than previously? Maybe aging leads to a >slowdown in learning ability or lessens the desire to study opponents's games. >Maybe a GM fades sometimes not because his powers have become weaker, but >because his opponents have (on the average) become smarter, better players, >maybe due to use of learning tools the older GM does not rely on, such as >computers, etc. A GM can lose his ability. Is a senile GM still a GM? In title, yes, in ability no. > E. If GMs learn about the weaknesses of a computer, and thereafter achieve >better results against it, perhaps the computer was and is still a GM in >strength, but the competition has improved itself (relative to the program). If >the GM humans become better players (even if only relative to one program), then >the computer should not have to win as great a percentage of the points to still >be considered GM in strength. I hear arguments about the ability of the human >GMs to learn and improve their play against the computers, but no corresponding >credit is then given to the computers for playing against stronger (more >learned) GM opponents. Boxers. One guy charges like a bull, but can't box worth a darn. So his opponents learn to sidestep him and keep him at arm's distance instead of going toe to toe. Instead of a world contender he becomes a has been. But he was never world-contender material, it only looked like it because nobody knew about his weaknesses. >4. Consider this--I compile a poly-program composed of all the program versions >that have played at tournament time controls against human GMs. I run this >poly-program on a poly-computer composed of all the hardware circuits and >processesors that those programs have played on against human GMs in the past. > >Assume this poly-program randomly utilizes only some selected combination of >program and processor that had been utilized in the past? > >Is my poly-program, running on a poly-computer, a GM (let's say, over 2500 ELO >TPR), if its predecessor parts and predecessor programs in individual >combinations achieved an average of 2500+ ELO TPR (calculated on the whole)? Build it first, then I'll answer your question. >Is my program a GM in style, because it changes its style of play flexibly and >often enough to make preparation by the opponent a very difficult matter? > >These are simply some things to ponder--I am not drawing any conclusions as yet. Good. Neither am I. That's my whole point.
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.