Author: Robert Hyatt
Date: 18:20:18 10/19/00
Go up one level in this thread
On October 18, 2000 at 18:11:41, Ratko V Tomic wrote: >>You miss the point. We _know_ it will sacrifice material. In positions >>where it is not clear that it wins. IE the Rc6 move. > >Well, you had put the same essential objection in several different ways (and >not only in this thread), and I had picked above the instance with the most >objectionable expression of your core issue with GT (which was not the form you >expressed above). > >What it comes down to is that you don't believe anyone can come up with a >sufficiently reliable evaluation/prediction of that type, at least not nearly as >reliable as the "piece value." I don't believe I said that at all. I simply pointed out that going too far is extremely dangerous. I don't have a clue whether Tiger loses games because of this, in significant numbers. I have certainly seen it lose some on ICC when it thought it was winning significantly, however. But it is more than possible that the big scores were not the problem and that it would have lost anyway... I tend to think "gambling a dollar might be ok because I can afford a dollar here and there." But gambling 1000 dollars is a different matter altogether. My wife would take issue there. :) >So your advice is, one should stick with small >adjustment style of positional evaluations (which also have the advantage of >keeping all such adjustments in the range of small perturbations, thus making >them mathematically nice, additive components). > >We are not dealing here with a formal system where one could prove to you one >way or the other by offering a chain of logical deductions from a commonly >accepted premises to the conclusions. The best one can do a priori is present a >collection of plausibility arguments and analogies. And what is plausible to me >is obviously not plausible to you. > >The only real proof is in the performance. And on that count GT seems to be well >on the way (lets see, among others, how it does in Dutch Open) to proving its >view. Another factor to take into account is that its current impressive >performance, the scores and the beta testers' judgments/impressions, is achieved >by a largely untuned (at least in its key novelty areas) new engine. Either >Christophe was very lucky to strike the right tuning values in the first few >shots (so it won't improve noticably in version 2), or he did come up with >something really new, an entirely new evaluation technique perhaps, which will >boost the state of the art by at least couple hundred points, while making games >much more enjoyable to humans, by the time it gets fully refined and extended to >the other aspects of the play (besides king attacks). > >In any case, it is unfair to classify it at this stage as just another gimmick, >some kind of cheap king safety evaluation tuning effect that many have toyed >with and failed with. I think it is already doing much better than what such >cheap "novelties" would have shown. Reading between the lines of Christophe's >posts here during the last year, I do see hints of a crystallization process of >a new technique (which ended up as GT), something which grew out of, or was >inspired by, some kind of cross between Lang's and Kittinger's ideas (at least >as he glimpsed at them through the reverse engineering of their programs). Since >Cristophe probably won't say openly here what exactly he did in GT, check his >posts since mid-1999 which mention Lang, Mephisto, Kittinger, Constelation, and >I think you will see (as I do) something new was cooking there. Maybe, maybe not. It is another "black box" for the moment. Once it is released, the lid will be off and details can be discovered pretty easily. Then we will know more. Right now, a racer has found a new profile for the lobes of his camshaft and is making a bit more horsepower than the rest. But just knowing this has happened is often enough to catch up to him. Same thing happens in computer chess each year...
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.