Author: Russell Reagan
Date: 11:26:27 05/20/02
Go up one level in this thread
On May 20, 2002 at 08:23:35, Eric Baum wrote: >How much do modern programs benefit from >developments beyond alpha-beta search +quiesence >search? So, if you did the same depth search, >same quiesence search, same opening book, >same endgame tables, but replaced the evaluation >function with something primitive-- say material >and not much else-- how many rating points would you >lose? > >My recollection is that one of the Deep Thought thesis >showed a minimal gain for Deep Thought from >extensive training of evaluation function-- >it gained some tens of rating points, but >less than it would have gained >from a ply of additional search. Has that changed? You speak of Deep Thought's "evaluation tuning", which as far as I can tell is nothing more than adjusting the weights of certain evaluation factors. For example, it might have 40% of it's score be material, and later on it adjusts that so that it uses 50% of it's total score from material evaluation. As far as adjusting the weights of an evaluation function, you aren't going to get incredible improvements. Take a simple example. If you had an evaluation function that was 50% material and 50% mobility, and you ran a "tuning" algorithm so that it would find optimal weights for each evaluation factor (material and mobility) then eventually the program would tune itself to better percentages for each factor. So it can help if your weights were bad to start with, but most people probably have at least an idea about how the weights should be set, so you will only see maybe rating increases of less than 100 points, which is still a good thing, but not drastic. Now, this is not the same thing as talking about how much a computer chess engine can improve it's knowledge and become a stronger player. For example, in the previous example, all you did was adjust the weights of the different factors. What you could do in addition to that (which would also make your program stronger) would be to add something completely new after you tune your weights for the previous factors. So instead of spending more time tuning the weights of evaluation factors, you would be better off adding in another piece of knowledge, such as maybe king safety or pawn structure information. Of course, I'm not saying that you should avoid tuning your evaluation factors' weights, but at some point you can't adjust them anymore. If you only had 2 factors (material and mobility) and you tuned them so that material accounted for 81% and mobility accounted for 19%, and you spent another year letting your program tune it's evaluation weights, and then after another year the program's weights were 81.2% for material and 18.8% for mobility, the program isn't going to play any better. Maybe a few rating points, and since the elo system is only an estimate of your strength anyway, a few points doesn't mean a thing. So, tuning the evaluation function's weights will help, but only a little bit. Adding new knowledge about chess will help a lot more once you have good values for your evaluation weights. Russell
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.