|
Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile RobotsBog'liq Alireza Fasih
5.2.3
Genetic algorithms
The ‘genetic algorithm’ concept was developed by John Holland in 1960s
[105]. It is an
effective method for determining the parameters for CNN. This method is inspired by the
mechanism of natural selection and genetics. It has been effectively used for solving
difficult search, optimization and machine-learning problems. It works by creating
genotypes (set of chromosomes) that represent the possible randomly chosen solutions.
The search evolves to improve the quality of chromosomes until the best chromosome is
found that represents the optimal solution[94]. The process of evolution occurs in the form
of generations and in each generation better chromosomes are sorted out. The parameters
of genetic algorithm that play an important role in the process of evolution and of finding
the best solution are the following: initial population, selection, reproduction, crossover,
mutation, and fitness function [94].
5.2.4
Initial population for the genetic algorithm
An initial random population of the chromosomes is generated. Each chromosome
represents a possible problem solution. All chromosomes are composed of a fixed number
of genes. The approach we used in genetic algorithms represents the chromosomes in the
form of real numbers instead of binary digits. This creates chromosomes of relatively
smaller sizes and the repeated operations of encoding and decoding are eliminated. For a
3×3 CNN, the total numbers of genes contained in a chromosome are 19. Thus, 9 genes are
for the control template; 9 for the feedback template and 1 for the bias. Similarly, if used
for 5×5 CNN, the total number of genes in a chromosome would be 51.
|
| |
