|
Abstract:
|
Evolutionary Algorithms (EAs) use a simplified
abstraction of biological evolution to interact with a fitness landscape
over multiple generations. The traditional approach to
exploring evolutionary motivated search relies on performance
comparisons of competing designs for a common problem domain.
This approach has been useful in developing an intuition
of when one mechanism is superior to another. However, this
intuition has developed in the absence of a clear understanding
of how fitness landscape topology impacts the search process in
high-dimensional space, due to the lack of high-dimensional
visualization tools. Proposed is a behavior measure derived
from a known set of all problem domain optima, which is used
as system of landmarks. Using these landmarks, evolutionary
system progress can be tracked in the problem domain and
characterized with an information theoretic metric. Thus, this
technique provides an intuition that takes into account the
problem domain topology, allowing the behavior of different
algorithm configurations to be compared as they interact with
a given topology. |