Author: Ratko V Tomic
Date: 10:02:57 10/21/00
Go up one level in this thread
This is a highly representative situation which shows the key achiles heel present day programs have against humans -- they lack the simplest common sense reasoning, most importatly the reasoning involving detection of simple invariant features of a position, the features which remain constant from move to move, as long as one side wishes so and no matter what the other side does. So far the CC technology has discovered only couple instances of this capability, the null move and the killer move heuristics, and CC had gained greatly from them, even though these are just two special cases among probably thousands of invariants possible. While one may try dismissing the problem by pointing out that Crafty can solve the problem under tournament time control, in fact if this were a leaf node in its search, or anywhere few levels down the tree, it wouldn't have had 90 seconds to solve it but maybe only 90 microseconds, and surely miss it. And on a less drastic scale, the same lack of simple common sense reasoning (about invariant features of a position), occurs all the time, in any game, be it in the actual position or in the millions of positions in the search tree. It wouldn't actually be that hard to write a special code to detect this particular kind of situation, but one couldn't afford to run that kind of check on every leaf node. But, if one were to check it only along the PV in between the iterations, one could run dozens or even hundreds of such specialized detectors. The results of these detectors of invariants could then trigger the matching full invariant analysis routines which would establish the particular invariant feature (or its absence) with absolute certainty. (The initial quick detectors are triggered when a given feature doesn't change its value as one descends along the current PV, but they don't prove the full invariance.) In the Uri's example, one of the quick detectors would trigger since the PV at, say, depth 4 would contain only checks whenever white has turn (a feature value which doesn't change along PV) and their avoidance through king movements. A specialized algorithm would then kick in to verify that white indeed cannot escape perpetual check here. The commonly quoted estimate that top players use 50,000 "patterns" in fact likely means the 50,000 of these kind of specialized quick detectors which can trigger the more elaborate matching verifiers. That's surely a long way from the couple (or a handfull) that programs currently use. While most of the 50,000 are likely well below the consciousness threshold, there are probably hundreds which could be retrieved in interviews (or data-mined out of chess theory books).
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.