Illinois Genetic Algorithms Laboratory

Srivastava R. P. (2002)
TR No.: 2002021 | Download PDF | Download PS

Abstract:

Given a sufficiently large population size and run duration, genetic lgorithms yield satisfactory results for a range of problems. However,
when resources and time are limited, as is the case with any real-world
problem, it is imperative to use those resources judiciously.

This master’s thesis studies the behavior and performance of genetic
algorithms under the influence of multiple-epoch runs versus a single
epoch run for limited resources or function evaluations. It includes
study of the temporal supply of building blocks and the issue of
population size involved therein. The thesis promises a key turn to the
much talked about battle between crossover and mutation operators of
GAs. The results end the question of choosing the better of the two
operators by showing that they actually share a symbiotic relationship.

This time continuation study aims at arriving at a prescriptive design
for efficient genetic algorithms. It aims at maximizing quality in a
given run time or minimizes run time for a given solution quality by
balancing parallel processing (a single large population) versus serial
processing (several small convergence epochs).

Methods adopted include empirical and theoretical analysis. The study
progresses by creating a set of abstract idealized and realizable
continuation operators, analyzes them and considers computational
approximations of them. Connections between the abstract operators and
theory and real GAs is explored with various continuation operators.