IEEE Computational Intelligence Magazine - November 2022 - 56

exploited to improve the learning efficiency
and generalization performance
ofeach task-specific model [23]. Owing
to these merits, multitasking learning
has been successfully applied to different
fields such as natural language processing
[24], speech recognition [25], computer
vision [26], drug discovery [27]
and so on.
Recently, evolutionary multitasking
(EMT), as an emerging sub-field of
multitasking learning, has attracted
much focus. Specifically, EMT [28],
[29], [30] aims to solve multiple optimization
tasks simultaneously via evolutionary
knowledge transfer. In the
multitasking setting with K tasks, the kth
task, denoted by Tk, has an objective
function fk with the search space Xk.
EMT algorithms are to find a set of solutions
fx1; ...; xKg, such that
x
k ¼ arg min
x2Sk
where x
fkðxÞ; k ¼ 1; ...; K (1)
k is the optimal solution of Tk,
and Sk is the decision space ofTk. EMT
algorithms can search multiple decision
spaces simultaneously via implicit parallel
ofthe population-based optimization
algorithms.
Gupta et al. [28] proposed the first
evolutionary multitasking algorithm
named MFEA, whose inspiration was
from the biocultural models ofmultifactorial
inheritance. In MFEA, the biological
and cultural building blocks (genes
and memes) were transferred from
parents to their offspring. In addition, a
cross-domain optimization platform that
allowed one to solve diverse problems
concurrently was also developed. The
numerical experiments on continuous,
discrete, and a mixture ofcontinuous and
discrete optimization problems demonstrated
the effectiveness and the efficiency
of the proposed MFEA. In recognizing
the superiority of MFEA, more evolutionary
multitasking algorithms have
been suggested. For example, inspired by
MFEA, Bali et al. [31] proposed a novel
multifactorial evolutionary algorithm
with online transfer parameter estimation,
named MFEA-II. Compared with
MFEA, the proposed MFEA-II had at
least two advantages. 1) Although the
mode of transfer remained the same as
MFEA, the parameter inducing the
extent of transfers was modified to take
the form of a symmetric matrix. This
allowed effective multitasking across
more than two tasks with possibly diverse
intertask relationships. 2) The transfer
parameter matrix was continuously
learned and adapted during the course of
the multitasking search. A series of synthetic
benchmarks as well as a practical
study validated the efficacy ofMFEA-II.
Bali et al. [32] further designed a cognizant
evolutionary multitasking approach
called MO-MFEA-II, whose aim was to
analyze how similar concepts can be
applied to the domain ofmulti-objective
optimization (MOO). Specifically, the
MO-MFEA-II utilized probabilistic
modeling to capture intertask relationships
amongMOO tasks, with which the
designed approach could be aware of
when and howmuch knowledge is to be
transferred. Wu et al. [17] suggested a
multitasking genetic algorithm (termed
MTGA), where the biases among different
tasks were estimated and utilized for
multitasking optimization. Specifically,
in MGTA, the biases between two tasks
were firstly estimated by calculating the
means of some of their fittest chromosomes.
Then the biases were deleted
from the chromosome transfer, as a result,
the optimal solutions of these two tasks
were getting closer, and more effective
knowledge transfer could be performed.
The experimental results on nine benchmarks
demonstrated the effectiveness and
the efficiency of MTGA. Furthermore,
the authors applied their suggested algorithm
to the fuzzy system optimization,
where a novel optimization strategy was
proposed for fuzzy system design. Empirical
studies on the fuzzy system logic controllersjustified
the superiority ofMTGA
over the state ofthe art. Unlike the above
works that conduct knowledge transfer
across tasks implicitly, in [33] Feng et al.
developed an explicit evolutionary multitasking
algorithm (EEMTA) for combinatorial
optimization problems, where
knowledge was transferred explicitly
among tasks in the form oftask solutions.
The case study on an illustrative combinatorial
optimization problem, such as
capacitated vehicle routing, showed the
promising performance ofEEMTA.
Following EEMTA, Tang et al. [34]
proposed a novel and efficient explicit
56 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2022
intertask transfer method, where an
intertask knowledge transfer strategy
was implemented in the low-dimensional
subspaces via a learnable alignment
matrix. In addition, to balance the
fidelity and the robustness in the intertask
optimization process, the aligned
subspaces and the original search spaces
were encompassed in the proposed
method. Comprehensive experiments
on the synthetic and practical benchmark
problems exhibited the superior
performance of the proposed method.
Meanwhile, Du et al. [35] suggested a
knowledge transfer-based evolutionary
algorithm (KTNEA) for multimodal
optimization by mining and exploiting
modal similarities. In KTNEA, a centeraligned
normalization strategy (CANS)
was firstly designed to map all of the
sub-populations corresponding to identified
modals into a unified 0-1 space,
and then a distribution similarity-based
transfer strategy (DSTS) was developed
to guide knowledge transfer among subpopulations.
The experimental results
on CEC'2013 niching benchmark suite
indicated that KTNEA performed competitively
in comparison with the state of
the art. Hao et al. [36] combined the
evolutionary multitasking learning with
hyper-heuristics and proposed a unified
framework of graph-based evolutionary
multitasking hyper-heuristic (EMHH).
The experimental results demonstrated
effectiveness, efficiency, and increasing
generality of EMHH, compared with
single-tasking hyper-heuristics.
Recently, Chen et al. [37], [38] proposed
two works that combined the evolutionary
multitasking learning with
feature selection in high-dimensional data
classification. In their first work [37], to
tackle the issues of local optima and
expensive computation in high-dimensional
feature selection, a new PSO-based
algorithm (named PSO-EMT) was proposed,
where the multifactorial optimization
was adopted as the evolutionary
multitasking paradigm. In PSO-EMT,
two related tasks about the target concept
were firstly generated by considering the
importance offeatures. Then, a new crossover
operator was utilized to share information
between these two related tasks.
Moreover, a variable-range mechanism
and a subset updating mechanism were
IEEE Computational Intelligence Magazine - November 2022

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