Author: Moritz Berger
Date: 11:25:45 10/03/98
Go up one level in this thread
On October 03, 1998 at 12:32:26, Robert Hyatt wrote: >this won't change a thing. After twenty games against each, whether the >games are played consecutively or alternating opponents, the learning will >be exactly the same, on a per-opponent basis... What you would see is >that fritz would have a lower score for the first half of the games, still, >and it would have a better score for the second half as the learning takes >effect. But intermixing opponents won't change a thing... I think that "intermixing opponents" might change the performance insofar as Fritz tries to hit the same weak spot over and over again as long as the opponents allows. Imagine a program like M-Chess which tries to repeat even lost lines and tries to find improvements towards the end of its book line and beyond with positional learning in some cases (I guess this behaviour depends upon evals during the previous game which M-Chess lost). Fritz might win several times in a row by staying within the same ECO system. If Fritz had won against M-Chess, tried the same line against Junior 5 (maybe Fritz lost), and then meets M-Chess again, it will try to chose a different type of opening first before stumbling upon this weak spot again. In the long run you are perfectly right that lifetime learning isn't influenced by "intermixing". That's the main reason why I completely disagree with Uri: Having no learning at all (PowerBook on CD) or having learning data from 2000 games against 15 different opponents makes a HUGE difference, even if playing against "unknown" opponents. In this case, it doesn't matter if opponents had played in "matches" or in "intermixed", isolated games. Moritz
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.