Self-adaptive learning for hybrid genetic algorithms

Local search can be introduced into genetic algorithms to create a hybrid, but any improvement in performance is dependent on the learning mechanism. In the Lamarckian model, a candidate solution is replaced by a fitter neighbour if one is found by local search. In the Baldwinian model, the original...

Full description

Saved in:
Bibliographic Details
Published in:Evolutionary intelligence 2021-12, Vol.14 (4), p.1565-1579
Main Authors: El-Mihoub, Tarek A., Hopgood, Adrian A., Nolle, Lars
Format: Article
Language:English
Subjects:
Applications of Mathematics
Artificial Intelligence
Bioinformatics
Control
Engineering
Fitness
Genetic algorithms
Machine learning
Mathematical and Computational Engineering
Mechatronics
Optimal control
Process parameters
Research Paper
Robotics
Searching
Statistical Physics and Dynamical Systems
Topology
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Local search can be introduced into genetic algorithms to create a hybrid, but any improvement in performance is dependent on the learning mechanism. In the Lamarckian model, a candidate solution is replaced by a fitter neighbour if one is found by local search. In the Baldwinian model, the original solution is retained but with an upgraded fitness if a fitter solution is found in the local search space. The effectiveness of using either model or a variable proportion of the two within a hybrid genetic algorithm is affected by the topology of the fitness function and the details of the hybrid algorithm. This paper investigates an intelligent adaptive approach to decide on the learning mechanism to be used by an individual over the course of the search. Evolution is used to self-adapt both the frequency of a steepest-descent local search and the relative proportions of Lamarckian and Baldwinian inheritance. Experiments have shown that this form of adaptive learning can improve the ability to find high-quality solutions and can accelerate the hybrid search without the need to find optimal control parameters for the learning process.
ISSN:1864-5909
1864-5917
DOI:10.1007/s12065-020-00425-5