|
USENET News comp.ai.genetic
Säie: Combining fitnessesSeuraava säie: Quake Genetics Project! [Muut säikeet] [Muut uutisryhmät]
From: Marko Gronroos <magi AT magi PISTE yok PISTE utu PISTE fi>
Newsgroups: comp.ai.genetic
Subject: Combining fitnesses
Date: 19 Oct 1997 12:39:07 +0300
How should the combining of different fitness factors be done? Most
people seem to use simple weighting:
F_tot = w1*f1 + w2*f2 + ...
, where the weights sum to 1. Some people have referred to an MDL
(minimum description length) principle in information theory. I'm not
familiar with this concept, but would like to hear why it is relevant.
It seems to me that this way of weighting is sensible only when the
selection function is fitness-proportional, but people seem to be
using it also with rank-based selection. My not-very-experienced
intuition suggests that when using rank-based selection, the weighting
should be done like:
F_tot = w1*rank1(f1) + w2*rank2(f2) + ...
, where rankN(fN) is a ranking function for each fitness
factor. After that the combined fitnesses should be sorted again for
selection.
My particular problem is the evolution of neural networks where you
have to combine fitness factors like MSE of a network and some
complexity factors of the network.
Have I missed some point? References are welcome.
----
-- Marko Grönroos, magi AT utu PISTE fi (http://www.utu.fi/~magi/)
-- Genetic algorithm researchers do it with the fittest individuals --
----
|
Seuraava säie: Quake Genetics Project! [Muut säikeet] [Muut uutisryhmät] |