IEEE Computational Intelligence Magazine - May 2022 - 97

Selecting which data and determining
the weight (importance) of each
selected data for formulating the problem
model are two critical issues in data
fusion, which directly affect whether the
model can accurately describe the optimization
problem in the real world. EC
algorithms are promising methods to
address these two issues in data fusion.
There are two potential solution frameworks:
offline and online. In the offline
framework, EC algorithms use historical
data for data fusion to generate a static
problem model. The advantage of the
offline framework is that the EC algorithms
only need to execute once.
However, because the generated model
is static, the EC algorithms must generate
a robust problem model that is capable
in various situations. In the online
framework, a rough problem model is
formulated initially. Then, the dynamic
transportation data collected from the
environment are used by EC algorithms
to tune the problem model. The advantage
of the online framework is that it
can adapt to various situations. However,
it tends to yield overfitting of the problem
model and requires a real-time
response of the EC algorithms.
B. Higher-Accuracy EC
The optimization problems of intelligent
transportation in smart cities will
be more and more complex in the
future. On the one hand, many complex
features exist in the optimization problems,
such as large-scale data, multiple
optimal solutions, multiple optimization
objectives, multiple constraints,
dynamic environments, and expensive
evaluations. On the other hand, an
optimization problem is usually with
several complex features simultaneously.
For example, optimization problems
in traffic management will involve
large-scale data and dynamic environments
simultaneously.
Therefore, besides the traditional EC
algorithms, the recently-proposed EC
algorithms that have shown great performance
in solving complex optimization
problems like large-scale optimization
problems, multimodal optimization problems,
multi-/many-objective optimization
problems, and in solving more complex
problems like large-scale multiobjective
optimization problems [184]-[186] and
large-scale multimodal multiobjective
optimization problems [187] are of great
potential to be applied to solve the optimization
problems of intelligent transportation
in smart cities. Moreover, designing
efficient EC algorithms that can deal with
complex features and generate a highaccuracy
solution is a meaningful research
direction. There are three guidelines for
the design of such efficient EC algorithms.
Firstly, prior knowledge of the
optimization problems should be utilized
to provide search guidance in EC algorithms.
Secondly, specific search strategies
or evolutionary operators that adapt to
the problem features and maintain sufficient
search diversity for avoiding premature
convergence can be designed. Thirdly,
hybrid algorithms that benefit from the
advantage of different EC algorithms are
also a promising approach to enhance the
algorithm performance.
C. Real-Time EC
Many optimization problems in smart
cities require rapid response. For example,
traffic flow is highly dynamic, and only
real-time decision-making can achieve
good results. As another example, when
an emergency occurs, such as a chemical
plant explosion, the response must be
determined quickly to reduce casualties
and to prevent the situation from deteriorating.
However, the large-scale data in
the optimization problems presents challenges
for the design of EC algorithms to
achieve a real-time response.
The combination of EC algorithms
and distributed computing technology
is a potential solution [188]. In EC
algorithms, each individual in a population
can evolve in parallel. Thus,
deploying EC algorithms on distributed
computing platforms, particularly cloud
platforms, can greatly enhance time
efficiency [189]. However, the communication
(information sharing) among
individuals of EC algorithms in distributed
computing scenario is time-consuming.
Researchers can focus on optimizing
the population topology of EC
algorithms to reduce communication
among individuals while maintaining
search efficiency. In addition, resource
allocation is another research topic. The
charging of cloud computing resources
follows the " pay-as-you-go " principle
that calculates charges based on resource
usage. Thus, time and cost efficiency
should be considered simultaneously.
Researchers can focus on the formulation
of resource allocation schemes for
efficiently utilizing computing resources
(e.g., maximizing algorithm performance
while minimizing computing
resources). Moreover, the latest and
modern matrix-based EC [190], generation-level
parallel EC [191], scale-adaptive
fitness evaluation EC [75], and
resource-aware distributed EC [192]
have great potential to reduce execution
time in achieving real-time response.
D. Data-Driven EC Algorithms
When designing EC algorithms to solve
the optimization problems of intelligent
transportation in smart cities, the related
data can be analyzed to generate search
guidance or knowledge for driving the
evolution of EC algorithms. Machine
learning (ML) techniques such as the artificial
neural network and support vector
machine are promising approaches for
data analysis, which have been widely
researched in recent years and have shown
great performance. Therefore, the incorporation
of ML techniques for designing
an efficient data-driven EC algorithm is a
promising research direction.
There are two types of related data
that can be utilized. One is the problem
data (i.e., the data in the optimization
problems of intelligent transportation in
smart cities). For example, when solving
expensive optimization problems, the
evaluation of a candidate solution
involves massive computational tasks and
thus usually takes a very long time or
very high cost. ML techniques can analyze
the existing evaluation results of
some candidate solutions, and obtain a
surrogate model to evaluate the new
candidate solutions for saving time and
cost [193], [194]. The other type of
related data is the algorithm data, i.e.,
the data generated by the EC algorithms.
For example, when solving
MAY 2022 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 97
IEEE Computational Intelligence Magazine - May 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2022
