IEEE Computational Intelligence Magazine - February 2021 - 24

learning and discusses how TL can be integrated with EAs to
achieve better optimization performance.

240
Documents

220
200

A. Evolutionary Algorithm

180

EAs denote a class of population-based metaheuristic optimization algorithms that are inspired by biological evolution, such as
reproduction, mutation, and selection [1]. It has been applied to
solve many problems that cannot be easily addressed by traditional
approaches, such as NP-hard combinatorial optimization problems [20], optimization problems with a large number of decision variables [21], and optimization problems with conflicting
objectives [22]. Typically, upon deciding on an encoding representation, e.g., binary encoding [23], gray encoding [24], and real
number encoding [25], the algorithm starts with a randomly
generated initial population. After fitness evaluation of all individuals in the population, selection is performed to identify fitter
individuals to undergo reproduction for generating the offspring.
Then, genetic operations such as crossover and mutation will be
carried out [25]. Such an evolutionary process will be executed
iteratively until certain predefined stopping criteria are satisfied.
From the workflow of EAs, we can observe that the optimization process is based on the movement of individuals in a population within the search space of the problem of interest.
Therefore, the population in an evolutionary search contains key
information for solving the problem, and useful traits could be
learned from moving traces of the population and be transferred
across problem domains if applicable to guide the search towards
enhanced optimization performance.

160
140
120

19
20

18
20

17
20

100
Year
FIGURE 1 Number of published research papers (recorded in Scopus) related to ETO.

problem-solving [14]. As depicted in Fig. 1, by a simple search of
the terms " Evolutionary Transfer Optimization " and " Evolutionary Transfer Algorithm " in Scopus1, the number of published
research papers increases from approximately 110 in year 2011 to
approximately 220 in year 2019. Note that there may be publications within the scope of ETO that are missing in the statistics,
due to the simple criteria used in the search. The ETO aims to
improve the performance of EA solvers in terms of the quality of
solution and speed of search by learning and transferring useful
traits across related problems in the form of solutions, structured
knowledge, etc. In the literature, various ETO approaches have
been proposed to solve both benchmark and practical optimization problems, including dynamic optimization problems [15],
multi-objective optimization problems [16], multitask optimization problems [17], combinatorial optimization problems [18],
and expensive optimization problems [19]. However, to the best
of our knowledge, there is no or little effort made to discuss and
categorize existing ETO approaches. Moreover, it is observed that
existing ETO research progresses disjointedly, which may impede
the development of new methodologies and applications in this
emerging field. This paper thus presents an exposition of existing
ETO approaches according to the type of problems being solved.
It also highlights some challenges faced by the ETO and discusses
a number of future research directions. We hope this paper will
attract attention from researchers in both evolutionary computation and machine learning communities to further design
advanced ETO approaches to address the ever-increasing complexity of optimization problems.
The remainder of this paper is organized as follows. Section
II presents the background of ETO, which includes the introduction of evolutionary optimization and transfer learning.
Section III gives a review of existing state-of-the-art ETO
approaches. Discussions on future research directions of ETO
are given in Section IV. Conclusions are drawn in Section V.
II. Background

This section provides a brief review of EAs, including the definition and workflow of a generic EA. It also introduces transfer
1

https://www.scopus.com/

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2021

B. Transfer Learning

It is known that humans naturally transfer knowledge between
tasks, such as by recognizing and applying relevant knowledge
from previous problem-solving experiences when a new task is
encountered [26]. Usually the more relevant a new task is to
those already solved in the past, the easier the new task can be
solved efficiently. In the context of computer science, transfer
learning has been used to leverage the knowledge learned in
one or more source tasks to improve the learning performance
of a related target task. As depicted in Fig. 2, the tasks can be in
the area of traditional machine learning, such as classification,
regression, reinforcement learning, and deep learning [27], [28].
The knowledge learned and transferred across tasks can come
from diverse representations [29], e.g., neural knowledge, matrix
knowledge, and tree knowledge, which is often problem dependent. TL has attracted increasing attention in recent years, and
many algorithms have been proposed to improve the learning
performance of TL in various applications, such as face recognition, indoor localization, and deep learning training [30].
The success of TL in machine learning shows that it is
capable of reducing the effort needed to build a model from
scratch by using fundamental logic or structured knowledge
within one domain and applying it to another. To achieve better optimization performance, instead of starting an evolutionary search from scratch, TL can be integrated into EAs to
generalize what has been optimized in the past. However, the

IEEE Computational Intelligence Magazine - February 2021

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