Abstract
The notion of ``evolvability'' --- the ability of a
population to produce variants fitter than any yet
existing --- is developed as it applies to genetic
algorithms. A theoretical analysis of the dynamics of
genetic programming predicts the existence of a novel,
emergent selection phenomenon: the evolution of
evolvability. This is produced by the proliferation,
within programs, of blocks of code that have a higher
chance of increasing fitness when added to programs.
Selection can then come to mold the variational
aspects of the way evolved programs are represented. A
model of code proliferation within programs is analyzed
to illustrate this effect. The mathematical and
conceptual framework includes: the definition of
evolvability as a measure of performance for genetic
algorithms; application of Price's Covariance and
Selection Theorem to show how the fitness function,
representation, and genetic operators must interact to
produce evolvability --- namely, that genetic operators
produce offspring with fitnesses specifically
correlated with their parent's fitnesses; how blocks of
code emerge as a new level of replicator, proliferating
as a function of their ``constructional fitness'',
which is distinct from their schema fitness; and how
programs may change from innovative code to
conservative code as the populations mature. Several
new selection techniques and genetic operators are
proposed in order to give better control over the
evolution of evolvability and improved evolutionary
performance.
Copyright 1996 Lee Altenberg
Users
Please
log in to take part in the discussion (add own reviews or comments).