Author: Roy Eassa
Date: 13:46:42 04/17/02
Go up one level in this thread
On April 17, 2002 at 16:24:22, Uri Blass wrote: >On April 17, 2002 at 15:59:18, Roy Eassa wrote: > >>On April 17, 2002 at 15:48:49, Roy Eassa wrote: >> >>> >>>A given computer's rating will go down significantly (even though it does not >>>gain or lose one iota of strength objectively) if and when its human opponents >>>gain anti-computer skills. >>> >>>Does that make sense? >>> >>>I guess early ratings are one thing and ACTUAL STRENGTH is a different thing >>>that is much harder to measure (requiring much more scientifically controlled >>>circumstances). >>> >>>For humans versus humans, the two things (rating and actual strength) have >>>tradionally been closely related, except when the player is a young child who is >>>improving very rapidly. >>> >>>There is significant reason to believe that RATING and actual STRENGTH can get >>>*way* out of sync with each other when it comes to computers, due to the extreme >>>relevance of the anti-computer skills (and not normal chess skills) of the >>>humans they have faced. >> >> >> >>Also, most (nearly all?) computers that have gotten an early rating (using fixed >>hardware and software) have seen that rating drop SIGNIFICANTLY over time, as >>humans learn better how to play well against computers. >> >>Does that mean: >> >>a) The computer is getting steadily weaker at chess? or >> >>b) Humans are quickly getting much better at chess? or >> >>c) A computer's early rating is NOT an accurate reflection of the computer's >>actual chess strength, but is SKEWED by the fact that humans lack a special >>skill that is required in order for them to score accurately against computers >>-- a skill that is SEPARATE and distinct from the traditional skill most human >>chess players have focused on? >> >>d) Some other explaination (please fill in)? > >A computer with constant hardware and software should not be allowed to get a >rating against humans if it cannot change it's evaluation function and players >can repeat similiar strategies to beat it. > >If the evaluation function is not changed after learning from games then it >should not get a rating without changes in the software. > >Usually changing the evaluation is done by the programmers. >I think that it can also be done by automatic learning of the program from >games. > >I think that the program also need to be private in order to get a rating >because in other cases the player may buy the machine and repeat a game that the >machine even did not know about. > >Uri I was not actually referring to a *specific* human learning the weaknesses of a *specific* program, but rather to the likelihood that most GMs will improve their overall anti-computer abilities with time, and some will improve this ability by a LOT. Take a strong program running on a fast PC -- I'll agree it could rate 2700 in current GM matches -- and put the only copy of it in a vault for 6 years. Let no human play it or study it during that time. Let the best 3 anti-computer GMs of the year 2008 play matches against it. I think there's a very significant probability that that SAME program running on that SAME hardware would then achieve a rating of 2500. Did it get 200 points weaker sitting in the vault?
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.