IEEE Computational Intelligence Magazine - February 2021 - 23

©SHUTTERSTOCK.COM/ELLAGRIN

new generation of offspring iteratively, until a predefined
condition is satisfied. In contrast to exact mathematical optimization methods, e.g., linear programming [2], quadratic programming [3], and least squares [4], EA has strong search
capability in solving complex optimization problems that are
non-convex or multimodal in nature. Over the years, EA has
attracted much attention in both academia and industry, and
has been successfully applied to many applications, including
problems with single objective optimization [5], multi-objective optimization [6], and dynamic -optimization [7].
Despite the success enjoyed by EAs, it is worth noting that
existing EA solvers usually conduct the search process from
scratch, independent of how similar the new problem encountered is to those already solved in the past [8]. Therefore, existing EAs often do not learn from problems, and the search
capabilities of EA solvers do not grow with problem-solving
experiences automatically. However, in practice, problems seldom exist in isolation. Ignoring the search experience of previous optimization processes on related problems may lead to
unnecessary computational costs in repeated searches on similar
problems. The ability to generalize what has been learned
across problems is likely to have a great impact in solving the
ever-increasing complexity of real-world optimization problems. Keeping these in mind, the development of new learning

and transfer approaches for creating a new class of " learnable "
EAs is thus desirable.
In machine learning, the idea of taking advantage of useful
traits across problem domains towards enhanced learning performance has received significant interest under the term of transfer
learning (TL) in recent years [9], [10]. TL leverages the knowledge gained from source domains, where much high-quality
training data are available, to improve the learning model in a
related target domain containing fewer high-quality training data
[9]. However, existing research studies on TL have largely been
confined to machine learning applications, such as computer
vision [11], natural language processing [12], and speech recognition [13], where problem or application data are usually widely
available for knowledge learning and transfer. In the context of
evolutionary optimization, usually only a given objective function and limited problem-specific data are available a-priori, and
new and advanced knowledge learning and transfer approaches
are thus needed for the development of " learnable " EAs.
There is increasing interest in conducting research on evolutionary transfer optimization (ETO), which is a paradigm
that integrates EA solvers with knowledge learning and
transfer across related domains to achieve better optimization
efficiency and performance. It is noted that the notion of memetics is closely related to that of knowledge representations for

FEBRUARY 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

23
http://www.SHUTTERSTOCK.COM/ELLAGRIN
IEEE Computational Intelligence Magazine - February 2021

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