IEEE Computational Intelligence Magazine - August 2021 - 9

preceptors are a kind of regression model
that can directly predict the performance of
a neural network without any training and
is a hot research topic among the community.
However, the training of performance
predictors requires a large number of welltrained
neural networks, which is often
scarce in practice. To address this issue, this
paper first developed a new encoding strategy
of architectures for calculating the
graph edit distances among different architectures.
Also, two self-supervised learning
methods were designed to improve the prediction
performance by learning the meaningful
representations of neural architectures.
This can help to enhance the quality of the
training data fed to the performance predictors.
The experiments demonstrated
promising performance of the proposed
performance predictor against the peer
competitors. In addition, the proposed performance
predictor is integrated into an
ENAS algorithm for validation, and the
results also showed its superiority of searching
for promising neural architectures.
In the fourth paper, " Forecasting Wind
Speed Time Series Via Dendritic Neural
Regression " by J. Ji et al., proposes a
regressive version of the dendritic neuron
model (DNM), i.e., dendritic neural
regression (DNR), to forecast wind
power. Particularly, wind energy is one of
the fastest-growing green energy resources.
Precise forecasting wind power is crucial
in planning the power system and
operating the wind farm. However, since
the wind speed time series is with chaotic
properties and high volatility, traditional
methods are incapable of producing satisfactory
forecasts. DNM is a plausible biological
neural model and has the potential
to forecast wind power well. However,
DNM is originally designed for classification
problems. To this end, DNR is developed
based on DNM to forecast the
wind power that is a regression task. Like
other neural network-based models, the
performance of DNR is also voluntary to
its architecture design. To address this
problem, the states of matter search (SMS)
algorithm is used to search for the promising
architecture of DNR without much
manual effort. The experiments were
conducted on two benchmark datasets
with two different time intervals. The
results revealed that DNR can provide
superior performance compared to its
competitors, and DNR-SMS was an efficient
tool for wind speed prediction.
The fifth paper, " A Self-Adaptive
Mutation Neural Architecture Search
Algorithm Based on Blocks " by Y. Xue et
al., proposes the algorithm named SaMuNet
to tackle the problem of " loss of
experience " caused by ENAS algorithms
in their early search stage. Specifically,
most of the existing ENAS algorithms
mainly focused on investigating the search
space or evaluation strategy to enhance
their performance. However, the search
strategy also plays an important role during
the whole process. In addition, the
information of individuals in different
generations are also crucial to the performance
of the corresponding evolutionary
algorithm, which is ignored by most
existing ENAS algorithms. To address
both issues, SaMuNet incorporated a selfadaptive
mutation component into the
framework of the evolutionary algorithm
to effectively search for the neural architectures.
Furthermore, a semi-complete
binary competition selection strategy was
also designed into SaMuNet to prevent
population degradation and slow convergence.
In addition, motivated by the
recent advances of DNN models, the
building blocks of DenseNet and ResNet
were also designed as the search units of
SaMuNet. The performance of SaMuNet
was compared with 17 peer competitors
on CIFAR10 and CIFAR100 benchmark
datasets. The results demonstrated
that SaMuNet can outperform most of
them in terms of both the classification
accuracy and the consumed computational
resource.
The guest editors of this special issue
would like to thank Prof. Chuan-Kang
Ting, the Editor-in-Chief of IEEE Computational
Intelligence Magazine, for his
great support in initiating and developing
this special issue together. Many thanks to
all members of the editorial team for their
kind support during the editing process of
this special issue. Last but not least, we
would also like to thank the authors for
submitting their valuable research outcomes
as well as the reviewers who have
critically evaluated the papers. We sincerely
hope and expect that readers will find
this special issue useful.
Introducing IEEE Collabratec™
The premier networking and collaboration site
for technology professionals around the world.
IEEE Collabratec is a new, integrated online community
where IEEE members, researchers, authors, and
technology professionals with similar fields of interest
can network and collaborate, as well as create and
manage content all in one place!
Learn about IEEE Collabratec at
ieee-collabratec.ieee.org
http://ieee-collabratec.ieee.org
IEEE Computational Intelligence Magazine - August 2021

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