Author: Stuart Cracraft
Date: 08:43:21 09/05/04
Go up one level in this thread
On September 05, 2004 at 02:09:34, Uri Blass wrote: >On September 04, 2004 at 23:54:50, Stuart Cracraft wrote: > >>On September 04, 2004 at 18:40:28, Uri Blass wrote: >> >>>On September 04, 2004 at 17:35:43, Stuart Cracraft wrote: >>> >>>>On September 03, 2004 at 18:14:56, Uri Blass wrote: >>>> >>>>>On September 03, 2004 at 17:30:18, Stuart Cracraft wrote: >>>>> >>>>>>On September 03, 2004 at 16:52:34, Andrei Fortuna wrote: >>>>>> >>>>>>>On September 03, 2004 at 15:41:42, Stuart Cracraft wrote: >>>>>>> >>>>>>>>On September 03, 2004 at 05:08:01, Andrei Fortuna wrote: >>>>>>>> >>>>>>>>>This makes me think how funny would be if two engines play, engine A would have >>>>>>>>>all kinds of those extensions in case of check etc, engine B would have >>>>>>>>>implemented a good eval function (with many terms regarding positional play) and >>>>>>>>>in the match engine B leads engine A towards the positions where engine A >>>>>>>>>discovers those mate attacks and so forth ahead of engine B, but he is on the >>>>>>>>>losing side due to B's positional play. >>>>>>>> >>>>>>>>I think this kind of self-play event and auto-tuning and genetic algorithms >>>>>>>>in general are under-estimated by the computer chess programmers. Just >>>>>>>>because good results haven't been generated and there is no easy "elixer" >>>>>>>>doesn't mean we shouldn't be trying it. >>>>>>>> >>>>>>>>Think of the time-savings. Heck, your auto-tune doesn't have to produce >>>>>>>>Bob Hyatt hand-crafted Crafty evaluation coefficients for terms you have >>>>>>>>to find and prove first -- but even if you don't produce something other >>>>>>>>than what you are doing now but saving a lot of time, then you have profited >>>>>>>>more. >>>>>>> >>>>>>> >>>>>>>Hi Stuart, >>>>>>> >>>>>>>Wasn't talking about auto-tuning, just was thinking that if someone invests in >>>>>>>evaluation function versus someone who invests in various extensions - the >>>>>>>former wins the game. Of course in reality programmers usually take care of both >>>>>>>areas ... >>>>>>> >>>>>>>Andrei >>>>>> >>>>>>Yes -- I understand you weren't -- but there is a big savings if you do >>>>>>it right. >>>>>> >>>>>>For me, it is worth investigating as I don't want to spend the rest of >>>>>>my life tuning evaluation functions. >>>>> >>>>>I believe that I can earn more from adding new knowledge relative to tuning. >>>>> >>>>>Tuning can be done not automatically based on watching problems that repeat in >>>>>games. >>>>> >>>>>Uri >>>> >>>>If a problem repeats in games and the program loses, then tuning will >>>>try various things to prevent it. >>>> >>>>Look at Slate's "mouse" program and its learning capability. Highly >>>>effective yet simple. No need to even adjust coefficients. Just store >>>>a hash and a move in an avoid file. >>>> >>>>Imagine what tuning could do. >>>> >>>>I believe both Schaeffer and Marsland have very high expectations for >>>>the future of tuning via various methods. >>>> >>>>Stuart >>> >>>The problem is that there are things that you simply need to add new knowledge >>>if you want to fix them. >>> >>>It is not about changing parameters and I do not see how it can be done >>>automatically. >>> >>>Uri >> >>Absolutely concur that nothing, save neural, could discover new associations. >> >>But once you have identified a term in a linear or non-linear context, then >>the weight for it -- THAT is tunable. >> >>Certainly parameters cannot not easily be added from nowhere automatically. >>We programmers are needed for that. >> >>However, they can be dropped by auto-tuning with the evaluation function >>eventually either zeroing them out or in such a way after the auto-tuning >>to rebalance everything in a way that would zero out any terms that could >>be zeroed out, thus dropping non-essential knowledge. >> >>It is at least as complicated to setup something solid and general. >> >>Have you read Baxter et al and their KnightCap -- please explain that >>success story. >> >>From 1600 to 2500 with one blip on an opening book and probably some >>more blips on repeat wins by players playing the same moves over and >>over is an outstanding success story. >> >>Do I have it wrong??? Has anyone repeated their success????? >> >>Stuart > >I do not think that knightcap is a success story. >Movei is clearly better than Knightcap with no automatic tuning. > >Uri Uri, 1600 to 2500, which is documented on the chess server, is not a success story? You have a strange sense of the historical viewpoint! Why is 4 1/2 class increases with unattended (mostly) operation not significant? I think they curtailed their effort early and could have gone higher than 2500 with further autotuning. I disagree with you. Stuart
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.