IEEE Computational Intelligence Magazine - August 2021 - 74

kill the oldest individual from the population, thus avoiding
the premature convergence problem. Zhu et al. [51] combined
two methods while removing the oldest and worst
individuals. In addition, Xie et al. [28] used the roulette strategy
to assign probabilities to each individual. Individuals with
higher fitness values have a higher probability of survival. Sun
et al. [14], [29] used a binary competition strategy for environmental
selection.
In the traditional binary competition, two individuals are
randomly selected; the individual with lower fitness value is
eliminated. This means that if two good individuals are
selected, the relatively bad one will be killed. This process
can have a negative effect on the population. Therefore, this
paper proposes an environmental selection operator based on
a semi-complete binary competition. The implementation of
the method is shown as follows: First, a number k is
Algorithm 5 Self-adaptive mutation operator of SaMuNet.
Input: The maximal generation N, number of strategies N 3s
randomly generated, then k individuals with the highest fitness
values are selected from the parent population and the
offspring population. We retain the k individuals and use
the binary competition strategy to select the needed individuals.
In this way, it is ensured that the excellent individuals
in the population will not be eliminated. This
method, which incorporates the 'elite' [52] mechanism in
GAs, can effectively prevent the population from degenerating
during evolution.
Algorithm 6 shows the details of the proposed environmental
selection based on the semi-complete binary competition.
First, randomly generate a number k from (0, N/2], select
k best individuals from the parent population and the offspring
population, and then use the binary competition strategy to
select N - k from the remaining individuals. Finally, the new
population is obtained.
= ,
the mutation probability μ, the number of offspring q, i = 0, current
iteration (cIter), the number of iterations (Iter), flagIter = 0,
Ng = 5.
Output: The population of mutated offspring.
1: Put three operation {Adding, Removing, Replacing} into
strategy pool;
2: while i < Iter do
3:
4:
5:
6:
for j < q do
Randomly generate a number r in range from [0,1];
if r < μ then
pool to generate
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
else
nfflag 1 ;
is,
end if
jj ;1=+
end if
end for
SNN
FNN
gs
gs
# = sum of each column in nsflag ;,is
# = sum of each column in nfflag ;,is
Reset matrices nsflag ,is and nfflag ;,is
if cIter - flagIter = = Ng then
tempSNN
tempFNN
gs
gs
Update the
(tempSF ;h##+NN NNgs gs
gq#
end if
ii ;1=+
26: end while
27: Output the population of mutated offspring;
74 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2021
=
Algorithm 6 Environmental selection based on semi-complete
binary competition.
Input: The parent population Pt, the offspring population Qt, the
population size N.
Output: The next population Pt+1.
1:
PU qtt1+ !! ! {};
{} ,{},
# = sum of each column in SNN
# = sum of each column in FNN
gs#
gs#
;
probability: ()/##=
PtempS
Reset matrices SNN and FNN
flagIter = cIter;
gq#
;
1 NN Nsg s
;
2: Randomly generate a number k in range of (0,N/2];
3: while j < k do
4: p ! Select the k individuals with highest values from
PU ;tt
,
5: PP ;Ptt
++! ,
11
6: end while
7: for jk 1=+ to N do
8:
9:
10: PP ;qtt
++! ,
11
11: end for
12: Output the next population Pt+1;
P1, P2 ! Randomly selected two individuals from Ut;
q ! Select the one with higher fitness from Ut;
Choose one strategy with roulette from strategy
j
q ;
new
if the new offspring qnew
nsflag 1 ;
j
is,
=
is better than qj then
IV. Experiment Design
This section verifies the effectiveness of the CNN architecture
searched by our proposed algorithm on image classification
tasks. We first introduce the benchmark datasets (in Part A),
then introduce the comparison algorithms (in Part B), and
finally describe the parameter setting (in Part C).
A. Benchmark Datasets
CIFAR10 and CIFAR100 are color image datasets that are
close to universal objects. For state-of-the-art CNNs,
CIFAR10 and CIFAR100 are the most widely used benchmark
datasets [27]. Therefore, this paper uses CIFAR10
and CIFAR100 to verify the effectiveness of our proposed
algorithm.
CIFAR10 contains a total of 10 classes of RGB color images:
airplane, automobile, bird, cat, deer, dog, frog, horse, ship and
truck. The size of each image is 32 × 32. Each class has 6,000
images. CIFAR10 has a total of 50,000 training images and
IEEE Computational Intelligence Magazine - August 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - August 2021
