Author: Robert Hyatt
Date: 18:11:55 05/05/01
Go up one level in this thread
On May 05, 2001 at 20:04:06, Jesper Antonsson wrote: >On May 05, 2001 at 10:58:36, Robert Hyatt wrote: >>Don't forget that there are programs and there are programs. Ernst used >>Dark Thought. Monty and I used Crafty. Nothing says those two will behave >>the same way. IE a more speculative program will gain more from another ply >>(because it will help to make the speculative eval more accurate) than a >>conservative program will... > >True, but perhaps even more so for "bad" than "speculative". I would say that >the better the eval, the more diminishing gains you will see. This is quite >obvious when you think about it, if you get a correct best move early (which >will happen with better eval) the best move won't likely change after that. This >means that a steeper diminishing gains curve is a sign of quality. I think it is the opposite. The simpler the eval, the more pronounced the diminishing returns, and vice-versa. Deeper search with a good eval not only finds "tactical" things but it finds "positional tactics" as well... > In fact, the >diminishing gains curve could be used for evaluating eval/extensions or even >comparing different programs respective eval functions in an objective, >scientific way. One problem is that it takes a lot of time to conduct the >experiments, of course, but perhaps low ply searches would be good enough given >a better and larger test set, which would make the method somewhat feasible >anyway. :-) > >As an observation along those lines, I would say that it is likely that Craftys >eval were better than Dark Thoughts at the time the experiments were conducted, >since Crafty seems to have a somewhat steeper diminishing gains curve. >Some further calculations I've done indicate that at ply 14, DT may have the >correct move 18% of the time, while crafty is correct 22% of the time, which >would make your ply 14 as good as DT ply 15. Unfortunately, the number of >positions in the experiment were probably too few to conclude the above with any >real statistical significance, but it's an interesting thought nevertheless, >that such comparisons are possible. > >>Today we are doing about 12-14 plies. To get to 20 will require a machine >>roughly 3^6-3^8 times faster. We won't see that in 10 years, which was my >>point. For the forseeable future, the "diminishing returns" isn't going to >>be a factor. > >Yes, as I said, the gains taper off very slowly. On the other hand, for each >additional ply, trading speed for eval will be more productive, which is >something that can be considered a factor if you compare with the situation, >say, 15 years ago. If you consider just one or two years, however, I agree that >it isn't a big deal. > >Jesper
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.