The combination of genetic algorithm and neural network for
weight training consists of three major phases. The first phase is
to decide the representation of connection weights, i.e., whether
we use a binary strings form or directly use a real number form
to represent the connection weights. Since this paper uses a real
code genetic algorithm, what we have to do is just to set each neuron’s
connection weights and bias to its correspondent gene segments.
However, it is difficult to reach convergence using the
binary-encoded simple genetic algorithms (SGA) to solve the optimization
problems that have too much design variables. So, a real
code genetic algorithm was used to overcome the disadvantages of
SGA. The second step is the evaluation on the fitness of these connection
weights by constructing the corresponding neural network.
The objective function (shown in Eq. (3)) is selected as the
fitness function directly. Because of the generalization of ANN, its
model can be used as the knowledge source for the optimization
algorithm. This method can compute the objective function in real
time. The evaluation criterion of the individuals was ‘‘LOWS-BEST”.
The third one is applying the evolutionary process such as selection,
crossover, and mutation operations by a genetic algorithm
according to its fitness. The evolution stops when the fitness is
smaller than a predefined value.
The hybrid network learning process consists of two stages:
firstly employing GA to search for optimal or approximate optimal
connection weights and thresholds for the network, then using the
BP to adjust the final weights. The steps of learning optimal value
for network weights are achieved using the hybrid of GA–BP algorithm
as shown in Fig. 2. At first, the populations initialization is
done; then The fitness of every chromosome is evaluated by measuring
the value of the total mean square error, see Eqs. (1)–(3).
After evaluating all chromosomes, an intermediate population is
created by extracting chromosomes from the current population
using the reproduction (selection) operator. In this study, the roulette
wheel selection based on ranking algorithm was applied for
the reproduction operator. Chromosomes were selected in quantities
according to their relative fitness after ranking in the roulette
wheel operator and placement into the intermediate population.
Finally, the population of the next generation is formed by applying
the crossover and mutation operator to the chromosomes of the
intermediate population. Then the new chromosomes reproduced
by selection, crossover, and mutation operators are evaluated,