Illinois Genetic Algorithms Laboratory

Winward, P., Goldberg, D. E. (2006)
TR No.: 2006006 | Download PDF | Download PS

Abstract:
Previous work showed how fluctuating crosstalk in
a deterministic fitness function introduces noise
into genetic algorithms. Fluctuating crosstalk, or
nonlinear interactions of building blocks, were
modeled by higher-order Walsh coefficients. The
fluctuating crosstalk behaved like exogenous noise
and could be handled by increasing the population
size and run duration. This behavior held until the
strength of the crosstalk far exceeded the
underlying fitness variance by a certain factor
empirically observed. In this paper, we extend that
work by considering fluctuating crosstalk effects on
genetic algorithm scalability using smaller-order
Walsh coefficients on two extremes of building
block scaling: uniformly-scaled and exponentially-
scaled building blocks. Uniformly-scaled building
blocks prove to be more sensitive to fluctuating
crosstalk than do exponentially-scaled building
blocks in terms of function evaluations and run
duration but less sensitive to population sizing for
large building-block interactions. Our results also
have implications for the relative performance of
building-block-wise mutation over crossover.