IEEE Computational Intelligence Magazine - August 2021 - 77

C. Evolution Process on CIFAR10
To prove the effectiveness of the self-adaptive
mutation mechanism and semi-complete binary
competition strategy proposed in this paper,
the classification accuracy of the final searched
individuals is collected, and plot the statistical
results. In this paper, the population size is 20,
and the total number of generations is 20.
Figure 7 compares the proposed algorithm and GA. To
The simplest way to choose parents is to select the
individual with the highest fitness value from the
population each time, but this will cause the population
to lose diversity and fall into a local optimum.
intuitively show how the self-adaptive mutation mechanism
guides the evolution process, a tendency line is drawn by connecting
the averages over ten points before and after the current
point. It is noted that the accuracy of some individuals is
0% in Figure 7. Because the variable-length encoding strategy
is used, the length that the offspring individuals generated by
the crossover individuals is too long. Due to the memory limitation
of our experimental equipment, an out of memory
occurs. As shown in Figure 7, for SaMuNet, from the 1st generation
to the 7th generation, the average classification accuracy
rises faster. From the 12th generation, the average classification
accuracy is stable at about 90%, and the gap between individuals
narrows, which shows that the algorithm is gradually converging.
SaMuNet can achieve about 96% accuracy in the last
generation. For the comparison method GA, due to the lack of
guidance from the self-adaptive mechanism, the individual
accuracy rises slowly. After the 12th generation, the traditional
GA gradually converges too, but the final convergence accuracy
is only about 94%. The analysis proves that the self-adaptive
mechanism has better performance in the later stage, and it can
also guide the later evolution.
In addition, to prove the effectiveness of the semi-complete
binary competition strategy, this paper makes a comparison
between SaMuNet with semi-complete binary competition
and SaMuNet with the binary competition algorithm. The
conditions remain unchanged from the previous experiment. It
can be seen from Figure 8 that, compared with the binary
competition, the semi-complete binary competition strategy
appears to be more stable throughout the entire evolution process.
The competition method tends to retain bad individuals.
Especially after the 15th generation, there are some poor individuals
with an accuracy of 80%.
The self-adaptive mutation mechanism proposed in this
paper can effectively improve the search efficiency of the
algorithm, and guide the later evolution through the experience
accumulated in the early stage. In addition, the proposed
semi-complete binary competition strategy can
effectively maintain population stability and prevent population
degradation.
VI. Conclusions and Future Work
The goal of this paper is to design a CNN architecture search
algorithm by using SaMuNet with an adaptive mutation strategy
and a semi-complete binary competition strategy. This goal
is successfully achieved on CIFAR10 and CIFAR100. This
paper uses a block-based design method to accelerate the entire
CNN architecture design process, and uses an adaptive mutation
strategy to make the algorithm adaptively select mutation
strategies in different evolution stages, so that the algorithm can
guide the search process effectively. In addition, a semi-complete
binary competition strategy is also designed for environmental
selection. This strategy can better retain the elites. It can
be seen from the experimental results that our proposed algorithm
achieves better results in comparison with different
hand-crafted, semi-automatic, and completely automatic structural
design methods in terms of accuracy and consumption of
computing resources. SaMuNet still consumes a lot of computing
resources, thus, in the future, our team will be committed
to accelerating the process of evaluating individuals, and reducing
consumption of resources.
1.0
0.8
0.6
0.4
0.2
0.0
Tendency (SaMuNet)
Individual (SaMuNet)
Individual (GA)
Tendency (GA)
Generations
FIGURE 7 Scatter diagram of SaMuNet and GA.
1.0
0.8
0.6
0.4
0.2
0.0
SaMuNet With Semi-Complete
Binary Competition
SaMuNet With Binary Competition
Generations
FIGURE 8 Scatter diagram of semi-completely binary competition.
AUGUST 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 77
Fitness (Accuracy)
Fitness (Accuracy)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
IEEE Computational Intelligence Magazine - August 2021

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