Investigating hybrids of evolution and learning for real-parameter optimization
Abstract
In recent years, more and more advanced techniques have been developed in the field
of hybridizing of evolution and learning, this means that more applications with these techniques
can benefit from this progress. One example of these advanced techniques is the
Learnable Evolution Model (LEM), which adopts learning as a guide for the general evolutionary
search. Despite this trend and the progress in LEM, there are still many ideas and
attempts which deserve further investigations and tests. For this purpose, this thesis has
developed a number of new algorithms attempting to combine more learning algorithms
with evolution in different ways. With these developments, we expect to understand the
effects and relations between evolution and learning, and also achieve better performances
in solving complex problems.
The machine learning algorithms combined into the standard Genetic Algorithm (GA)
are the supervised learning method k-nearest-neighbors (KNN), the Entropy-Based Discretization
(ED) method, and the decision tree learning algorithm ID3. We test these algorithms
on various real-parameter function optimization problems, especially the functions
in the special session on CEC 2005 real-parameter function optimization. Additionally, a
medical cancer chemotherapy treatment problem is solved in this thesis by some of our
hybrid algorithms.
The performances of these algorithms are compared with standard genetic algorithms
and other well-known contemporary evolution and learning hybrid algorithms. Some of
them are the CovarianceMatrix Adaptation Evolution Strategies (CMAES), and variants of
the Estimation of Distribution Algorithms (EDA).
Some important results have been derived from our experiments on these developed algorithms.
Among them, we found that even some very simple learning methods hybridized
properly with evolution procedure can provide significant performance improvement; and
when more complex learning algorithms are incorporated with evolution, the resulting algorithms
are very promising and compete very well against the state of the art hybrid algorithms
both in well-defined real-parameter function optimization problems and a practical
evaluation-expensive problem.