Simulation of the Action Potential in the Neuron’s Membrane in Artificial Neural Networks - Advancing Artificial Intelligence Through Biological Process Applications - page 80

Figure 3. Modelling of the neuron's activation it is recurrent or not) from the input to the output

layer. In order to use a GA, first one must codify
the possible solutions as an array with n elements
or cells, which is going to represent a possible
solution, that is, an individual in the population
that represents a configuration of ANN.
In this case, each ANN has been codified with
two arrays. A weight array goes from the initial
layer to the output (Figure 5). The network connec-
tion weights, can be seen as a n * n matrix, where
n stands for the total number of neurons in the
network, so that if there is a connection between
i and j neurons, the weight of that connection will
be stored in the matrix intersection of row i with
column j . If it is a feedforward ANN, just a limited
number of weights are stored in the top part of the
matrix, while if the network is a recurrent one,
the matrix will be fully occupied. To represent the
time decreased activation of the neurons we need
another n * n matrix. This matrix is similar to the
previous one, but in this case, the values represents
the line gradients ( m ) of the formula of the Figure
4 (in this case the values will be always negative)
.The other array contains the activation functions
for each neuron in the network (Figure 6). This
array has two positions for each neuron: the first
one with the type of activation function (linear,
threshold, sigmoid, or tangent hyperbolic) and the
second one with the configuration parameters of
the chosen function for that neuron.
The ANN training process allows to choose
whether the network is recurrent or not, the param-
eters common to all the neurons in the network,
such as the value limit of the weights and the
recurrent network parameters, such as stage or
continuous training and the number of internal
iterations of the network before the output.
The GA to be applied uses an elitist strategy
of protection of the best individual against mu-
tation and substitution, and the codification of
the individuals is made through real numbers.
Once the ANN and GA general parameters have
been selected, the second phase is the creation
of a random population of individuals to start

Classical ANN Model

Instant of the
activation

Next cycle

0 1 0 1

0 1

Time Decreased Activation

0 m n 0 m n 0 m n

Instant of the
activation

m Instants
after

n Instants
after

0 m n

Time

been recently applied to the ANN field (Angeline
et al., 1994) (Ku et al., 1999) but it has just done a
good results to the training of different types of
ANN (calculating weights) and to the design of the
network architectures (number of layers, number
of neurons per layer and activation functions).

SySTEM DESCRIPTION

The first thing to use a GA is to codify the possible
solutions as an array of n elements. Each of these
solutions will be an individual of the population.
In this case, we can codify the ANN as an array
of gradients or weights (depending on the ANN if