Author: Stephen A. Boak
Date: 22:54:01 01/28/00
Go up one level in this thread
>On January 27, 2000 at 20:36:30, Dann Corbit wrote: <snip> >I suspect (without proof) that 1/2 of all computer program losses against other >computers is due to bad opening book advice. When you fall out of book at -200, >what can you do? You have an enormous disadvantage that most of the time will >prove fatal. In gambit openings, a human intentionally sacrifices a pawn or so for long term positional/tactical chances giving both opponents a seemingly lasting imbalance of material and positional features. Since a decent gambit line is not 'solved', it appears playable to the human mind mainly because the human can see the possibilities in the position, even when down the investment in material. The chances for the sacrificer may be based solely on the perceived, long lasting complications, which leave a rich field of choices available to both human players and a challenge to see who can better make plans and better navigate the tactical/positional minefields in the attempt to make his plan succeed. The computer can't really plan, but only examine tactics and make rather limited positional assessments within a limited search horizon. The computer assessments are made by relatively 'rigid' programming that isn't as flexible as the human mind at assessing changing positional and tactical features and weaving its way, on the fly, through intertwined imbalances. If one computer exits an opening down a pawn or two against another computer opponent, both silicon players will generally operate bereft of plans for some time thereafter. Neither will truly understand the position (from a long term strategic perspective), although either will see tactical/material mistakes and gains within its event horizon. The delicate dance of the creative and discriminating human mind (doesn't have to be a master!) among multi-faceted ideas (perhaps involving possible end-game advantage of outside passed pawn vs. middle game attacking chances upon the enemy king), based on the complex position on the board, will be impossible for the computer to emulate. As a result, the computer side that is down the material will more likely remain down in material during and after the planless shuffling that ensues. It will not have the human master's judgement regarding when to toss more material onto the sacrificial bier to fuel a hotter and more unstoppable attack, or when to bypass immediate material recovery to crank up the positional complexity to fever pitch and give the opponent a greater chance to go astray. Instead, each computer will manage its position rather stiffly, rather crudely, jockeying about to keep its material/positional score from lessening. The one will seek to maintain its material advantage, not realizing the critical moments when best play may require giving back the material in order to exchange it for a different plus--perhaps a lasting positional edge that will survive into the endgame (outside passed pawn, for example). The other will shuffle back and forth, sniffing for a tactical opportunity to win back material, even when it needs to find creative ways to give up even more material to best seek winning chances. When it finally sees a material gain in front of its silicon nose and within its silicon sniffing horizon, it will yelp like a starving dog and leap for the scrap, never realizing that it just threw away its long lasting positional pressure for a mere dry bone. Until programming improves enough to avoid the computer playing by some relatively simple animal-like instincts for survival based largely on relatively 'rigid' underlying material/positional tradeoff assessments--i.e. programmed evaluation scoring), and instead instills the skill of long term strategic planning, I fear the beast will continue to rut over meager scraps and dry bones that a GM would gladly feed it to lure the animal into a corner where it is ultimately trapped and destroyed. There have been several recent and instructive examples of such feral instinct problems by computers in comp-human play. I think GMs will have to continue to instruct the programmers, 1) by assisting them to tune their evaluations and better balance the oft-changing relative scoring for material and positional values, and 2) by defeating their programs in serious matches, showing 'en praxis' where are their Achilles heels (or silicon noses!). --Steve Boak
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.