Author: Robert Hyatt
Date: 12:14:26 11/09/00
Go up one level in this thread
On November 08, 2000 at 22:34:36, Bob Durrett wrote: >Chess engines [or programs] which can learn might be compared with regard to how >fast and how well they learn. One could define a chess-playing program's I.Q. >by creating a scale with the "average" learning computer arbitrarily assigned an >I.Q. of 100 and the smartest [fastest &/or best learning] currently available >program assigned an I.Q. of, say, 150. Similarly, one which learns poorly could >be assigned an I.Q. of, say, 50. A program without any learning ability at all >would be said to have an I.Q. of zero. The rest of the programs could be placed >on the scale in some manner which made sense. Just one brief comment. I don't think "learning" is really the right word, if you think of "learning" in the context of humans. You might beat me when I play the Sicilian as black, and I learn that I don't like any variations where you get to play f4 and g5 early on. So one loss can teach me to avoid a multitude of opening variations. A computer program is more like a rat in a maze, where it turns down the wrong path and gets shocked. It figures out that it doesn't like getting shocked, but the 'shock' is associated with a specific path. IE the rat won't learn that after turning down one red path and getting shocked, that maybe turning down red paths is bad. Programs do the same. And current "book learning" is really nothing more than 'electric shock therapy' to a computer. You can turn the learning scores way up and make it a cattle prod type of shock. Or you can turn them way down and make it a minor tingle... but the approach is the same, regardless of the voltage used. A human might get shocked and learn to (a) stay off red paths; (b) stay off paths that have odd numbers; (c) anything else that he might be able to recognize and generalize, to avoid getting a second shock. > >Chess-playing programs which can learn might differ in various ways, which might >make the definition of I.Q. somewhat awkward. For example, one program might >only be able to improve it's opening book. Another might be able to increase >it's strength in the middlegame, assuming that it plays worthy opponents who are >strong enough to teach the program something of value. Similarly in the >endgame. > >This raises several questions: > >(1) What are the I.Q.s of popular chess engines [or programs]? In what ways do >they differ with respect to their learning abilities? > >(2) Which engines/programs should be rated the highest with respect to it's >learning ability? [possibly broken down by phase of the game in which learning >can occur] > >(3) Are there any practical limits to how smart [high I.Q] chess >engines/programs can be made? What sets those limits? > >(4) If an off-the-shelf "high-I.Q." chess engine can be educated, then why not >deliberately educate it prior to delivery to the customers? The only >requirement seems to be the availability of qualified teachers. > >(5) Are there any practical limitations on how well educated a chess >engine/program can be? [Do their "brains" get "filled up"?] Could you have a >"PhD" engine who graduated, after several years of "study" from a "college" >where all of the "professors" were top GMs?
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.