IEEE Computational Intelligence Magazine - May 2022 - 50

Abstract-Until recently, the potential to transfer evolved skills
across distinct optimization problem instances (or tasks) was seldom
explored in evolutionary computation. The concept of
evolutionary multitasking (EMT) fills this gap. It unlocks a population's
implicit parallelism to jointly solve a set of tasks, hence
creating avenues for skills transfer between them. Despite it
being early days, the idea of EMT has begun to show promise
in a range of real-world applications. In the backdrop of recent
advances, the contribution of this paper is twofold. First, a
review of several application-oriented explorations of EMT in
the literature is presented; the works are assimilated into half a
dozen broad categories according to their respective application
domains. Each of these six categories elaborates fundamental
motivations to multitask, and contains a representative experimental
study (referred from the literature). Second, a set of recipes
is provided showing how problem formulations of general
interest, those that cut across different disciplines, could be
transformed in the new light of EMT. Our discussions emphasize
the many practical use-cases of EMT, and are intended to
spark future research towards crafting novel algorithms for realworld
deployment.
I. Introduction
O
ptimization is at the heart of problem-solving. Many
practical problems however possess non-convex, nondifferentiable,
or even black-box objectives and constraints
that lie outside the scope of traditional
mathematical methods. Evolutionary algorithms (EAs) provide a
gradient-free path to solve such complex optimization tasks, with
flexibility to cope with additional challenges such as expensiveto-evaluate
objectives [1], dynamics [2], etc. EAs are populationbased
methods inspired by Darwinian principles of natural
evolution, but, notably, fall short of simulating the phenomenon
in its entirety [3]. Unlike the tendency of natural evolution to
speciate or produce differently skilled sub-populations, the
update mechanisms of in silico EAs are usually crafted to evolve a
set of solutions for only a single target task. This naturally limits
the power of a population's implicit parallelism [4], often slowing
down convergence rates as useful skills from other related tasks are
not readily accessible. The concept of evolutionary multitasking
(EMT) addresses this limitation by offering a new perspective on
the potential of EAs.
The notion of generalizing beyond the ambit of just a single
task is set to mold the future of search and optimization algorithms,
especially since real-world problems seldom exist in isolation
[5], [6]. In scientific and engineering applications, for
example, building on known knowledge and existing solutions
can greatly reduce the time taken to explore and innovate new
designs-which could otherwise take days, weeks, or even
months to discover if probed in a tabula rasa manner [7]. Yet,
EAs continue to be crafted to work on problem instances independently,
ignoring useful information gleaned from the solving
of others. The notion of EMT fills this gap, launching the
inter-task transfer and adaptive reuse of information across distinct,
but possibly related, tasks. The transfer is achieved by
50 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2022
unlocking a population's implicit parallelism in a new class of
EAs equipped to jointly tackle multiple tasks.
EMT was put forward in [8], and has since attracted much
interest among evolutionary computation (EC) researchers. A
variety of algorithmic realizations have been proposed, including
the single-population multifactorial EA (MFEA) [8], multi-population
algorithms [9], or other co-evolutionary algorithms [10],
aiming for efficient solving of multiple tasks by maximally utilizing
mutual relationships through information transfer. To this end,
research questions in terms of what, how, and when to transfer arise
in the unique context of EMT. In what follows, a brief overview
of the ways in which today's EMT and transfer EAs address some
of these questions is provided. Since an in-depth methodological
analysis is not the focus of this paper, readers are also referred to
[11]-[13] for more comprehensive discussions on these topics.
Determining what to transfer emphasizes the type of information
unit and its computational representation [14]. Besides
genetic transfers of complete solution prototypes or subsets of
solution strings (e.g., frequent schema) [15], [16], other knowledge
representations have included probabilistic search distribution
models [14], search direction vectors [17], higher-order
heuristics [18], and surrogate models of expensive objective
functions [19]. Given the information type, how to transfer
becomes crucial when dealing with heterogeneous tasks (e.g.,
with differing search space dimensionality). Various solution representation
learning strategies for mapping tasks to a common
space have been proposed in this regard [20]-[24], with an
abstract categorization of associated strategies presented in [25].
Given what and how, discerning situations when to (or when
not to) transfer is a natural follow-up to maximize utilization of
inter-task relations while curbing harmful interactions. Increasing
efforts have thus been made to craft adaptive EMT algorithms
capable of online discovery of similarities even between
black-box optimization tasks. The gleaned similarity has then
been used to control on-the-fly the extent of transfer between
tasks [26], as opposed to earlier approaches that predefined and
fixed this quantity [8], [27].
Ongoing works in EMT are deeply focused on addressing
theoretical questions of the aforementioned kind, often assuming
synthetic settings with algorithmic tests run on idealized
benchmark functions. A mathematical proof of faster convergence
has also been derived under convexity conditions [28].
Given the methodological advances being made, the time is
deemed ripe to also draw the attention of researchers and practitioners
to the rich but nascent space of real-life applications of
EMT. From the design of multiphysics products [29] to social
network reconstruction [30], [31] and search-based software
optimization [32], EMT promises significant performance gains
in varied domains where multiple related problem instances
routinely occur/recur. Thus, with the goal of strengthening the
bridge between the theory and practice of EMT, this paper
makes the following twofold contribution.
❏ An extensive literature review from the perspective of the
real-world applicability of EMT is presented. Applicationoriented
explorations of multitasking are encapsulated in
IEEE Computational Intelligence Magazine - May 2022

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