androidtech - 10:17 pm on Feb 7, 2005 (gmt 0)

baron13,

Probably because there is no way to do a small run that would be representative of the full population, where the "population" is the billions of web pages that Google has. So anything they try in a microcosm (in the background), would have very little chance of scaling up to the real search universe because the keyword statistics of the microcosm population would have very little resemblance to the real search universe.

So instead, they probably run multiple tests in parallel, with different versions of the algorithm on different data centers. Analyze what works best, cross populate it to the other servers, and keep doing this until they get something they "like".

The behavior being seen is very much like that of doing a genetic algorithms run. If you want more info, just punch "genetic algorithms" into Google and start reading.

Wild speculation on my part of course.