Adaptive Genetic Algorithms - Computational Intelligence in Time Series Forecasting - page 320

9.5 Fuzzy Logic Controlled Genetic Algoithms
A number of scientists, after experimenting with probabilistic approaches for
improving GA performance, were not satisfied because, by pursuing this research
track, much vague and ill-structured knowledge and some highly exhaustive
computational procedures have to be used. They, therefore, started searching for
more comfortable and more efficient alternatives for solving this problem. To
escape from the probabilistic concepts and to by-pass the long-lasting calculations
they selected fuzzy logic as a possible tool for on-line adaptation of GA parameters
and for GA resources management. Lee and Takagi (1993) took this route in their
study and worked out a dynamically controlled genetic algorithm using a fuzzy
logic technique. Soon thereafter, Arnone et al. (1994) reported on fuzzy
government of a genetic population, and Bergmann et al. (1994) published their
experience with GA parameter adjustment using fuzzy control rules.
Dynamically controlled genetic algorithm is an algorithm that uses a fuzzy
knowledge-based system to control the GA parameters dynamically, mostly the
crossover, mutation rate, and the population size. In fact, it is a typical rule-based
expert system the inputs of which can be a combination made up of a genetic
algorithm and performance measures, such as the ratio of average to best fitness,
current population size or the mutation rate. The rules stored in the system reason
about the state of the measure values and recommend adequate actions. The
authors give a rule example: an increase in the present population causes the
sensitivity to mutation rate to decrease, along with the best mutation rate to use.
This can be programmed as follows:
IF the ratio of average fitness-to-best fitness is HIGH
THEN population size should INCREASE
IF the ratio of worst fitness-to-average fitness is LOW
THEN population size should DECREASE
IF mutation is SMALL and population is SMALL
THEN population size should INCREASE
The system developed was validated through a simulation example of an inverted
pendulum control, where it has shown much better behavioural results in pendulum
control than a GA with fixed parameters.
Government of the genetic population is a concept coined by Arnone et al.
(1994) for describing the process of on-line tuning GA parameters using a fuzzy
knowledge base. The concept is based on a fuzzy government module whose
inputs are statistical data periodically collected from the genetic algorithm and
whose outputs are the control parameters of the GA. In the concept, a facility is
embedded for monitoring the evolutionary process in order to avoid its possible
undesired behaviour.
Herrera and Lozano (1996) summarized the steps in building adaptive GAs
using fuzzy logic controllers as follows: