IEEE Computational Intelligence Magazine - May 2023 - 88

median accuracy values, respectively.
The extended lines at both ends of the
box represent the maximum and minimum
accuracies. The light blue boxes
show the results of Exp 1 5, and the
green box represents the results ofExp 6.
Among the light blue boxes, when
FIGURE 9. The impact of the length of the second stage on the three-stage EA.
C. The Effectiveness ofthe ThreeStage
EA and the Best Length
Division for the Three Stages
Six independent experiments are conducted
to investigate the effectiveness of
the three-stage EA. In each experiment,
an EA with the same number of evolution
rounds and different lengths for the
three stages is implemented. The classification
performance ofeach final population
is recorded. Five experiments,
named Exp 1 5, have the same fixed
length for the first stage but different
lengths for the second and third stages to
investigate the effects of the three-stage
length on the population. Another
experiment named Exp 6 only goes
through the second stage. In Exp 1 5,
the value ofG2 is changed to investigate
the impact of different stage lengths on
the validation accuracy ofthe final population.
The length ofthe second stage is
recorded. In Fig. 9, each box represents
the overall validation accuracy ofa population,
the length ofthe box represents
the accuracy deviations between individuals,
and the dot and dotted line in
the box represent the average and
the length ofthe second stage is set to 0,
the EA degenerates to the traditional EA
with fixed-criteria environment selection.
Fig. 9 clearly shows that the average
validation accuracy achieved by the traditional
EA is lower than the accuracies
ofthe other three-stage EAs. The second
stage helps the population to converge
to network architectures with better
classification performance since more of
the search space is explored.
When the length of the second
stage is increased, the population's
average accuracy initially increases and
then decreases later. When the length
of the second stage is set to 50, which
is implemented in Exp 6, the EA
degenerates to an aging evolution process.
Fig. 9 shows that the green box
and its extended lines have the longest
lengths, which indicates the most
extensive search instability; it illustrates
a prolonged second stage that causes
the population to generate more individuals
with poor fitness, reducing the
number of the optimal solution. The
box and extended lines of Exp 3 have
the shortest lengths, indicating that the
individuals in this population have the
smallest differences. Exp 3 shows that
utilizing a sufficient length for the third
stage improves the exploration ability
of the algorithm, which helps remove
poorly performing individuals and
increase the number of optimal solutions.
The stability of the search procedure
is also improved.
In summary, using an appropriate
FIGURE 10. Comparison of the three-stage EA, traditional EA, and random search method in
terms of the test accuracy and the number of parameters.
88 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2023
length for each stage can help to effectively
balance exploration and exploitation,
leading to better optimal solutions.
Based on the ablation experiments,
when G1 and G2 are set to 15 and 30,
respectively, the population shows the
best performance. The lengths of the
three stages are set to 15, 15, and 20, and
the values ofG1 ¼ 15 and G2 ¼ 30 are
used in the following experiments.
IEEE Computational Intelligence Magazine - May 2023

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2023
