IEEE Computational Intelligence Magazine - May 2022 - 52

... implicit genetic transfers naturally emerge
from standard evolutionary operations in unified
space, without having to craft ad hoc transfer
mechanisms.
value of wij
is increased to intensify the cross-sampling of solution
prototypes. In contrast, if solutions transferred from a given
source do not excel in the recipient target task, then the coefficient
corresponding to that source-target pair is gradually neutralized.
A complete algorithmic implementation of this general
idea is described in [14].
C. An Overview of EMT Methodologies
A plethora of EMT algorithms have been proposed lately.
Some of these either directly or indirectly make use of the formulation
in Eq. (3). Nevertheless, as noted in [11], most algorithms
typify one of the two methodological classes stated
below. An extensive analysis of these methods is not included
herein as excellent reviews are available elsewhere [12], [13];
only a handful of representative approaches are discussed.
(1) EMT with implicit transfer: In these methods, the
exchange of information between tasks occurs through evolutionary
crossover operators acting on candidate solutions of a
single population [37]-[39]. The population is encoded in a unified
space X using task-specific invertible mapping functions
}i .
Implicit genetic transfers materialize as solutions evolved
for different tasks crossover in X , hence exchanging learnt
skills coded in their genetic material. Over the years, a multitude
of evolutionary crossover operators have been developed,
each with their own biases. The success of implicit transfers
between any task pair thus depends on whether the chosen
crossover operator is able to reveal and exploit relationships
between their objective function landscapes. For example, in
[40], an offline measure of inter-task correlation was defined and
evaluated for parent-centric crossovers synergized (strictly)
with gradient-based local search updates. In [26], an online
measure was derived by means of a latent probability mixture
model, akin to Eq. (3); the mixture was shown to result from
the use of parent-centric operators in the single-population
MFEA. (Adapting the extent of transfer based on the coefficients
of the mixture model then led to the MFEA-II algorithm.)
Greater flexibility in operator selection can however be
achieved through self-adaptation strategies as proposed in [16],
where data generated during evolutionary search is used for
online identification of effective crossover operators for transfer.
(2) EMT with explicit transfer: Here, information exchanges
take place between multiple populations. Each population corresponds
to a task in MTO and evolves in problem-specific
search space ,.iXi
6 The populations mostly evolve independently,
between periodic stages of knowledge transfer. An
explicit transfer mechanism is triggered whenever a predefined
condition, e.g., transfer interval, is met [27]. For homogeneous
52 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2022
cases where X XX ,K
12
==g=
islandmodel
EAs for multitasking have been proposed
[41], with added functionality to
control the frequency and quantity of solution
cross-sampling [42]. Under heterogeneous
search spaces, mapping functions
XX ,ij
}ij :
"
for all ij ,! must be defined to reconcile
ith and jth populations. To this end,
while most existing EMT methods have made use of linear
mapping functions [27], [43], the applicability of fast yet
expressive nonlinear maps, as proposed for sequential transfers in
[25], [44], are deemed worthy of future exploration.
Both methodological classes have their merits. In the spirit
of the no free lunch theorem [45], algorithmic design preferences
must therefore be guided by the attributes of the application
at hand. Note that implicit genetic transfers naturally
emerge from standard evolutionary operations in unified space,
without having to craft ad hoc transfer mechanisms; hence,
implementation is relatively simple and scales well for large K.
However, composing a unified space and search operators for
heterogeneous tasks becomes highly non-trivial (operators that
work well for one task may not be effective for another). In
contrast, the multi-population approach of explicit transfer suppresses
the need for unification, allowing each task to hold specialized
search operators. But additional complexity may be
introduced in having to define ()KO 2
inter-task solution mappings
( ij} 's) under heterogeneous search spaces [13].
III. A Review of EMT in Action in Real-World Problems
The aim of this section is to draw the attention of both
researchers and practitioners to the many practical use-cases of
EMT. Prior literature exploring real-world applications is
encapsulated in six broad categories, together with representative
case studies and published results that showcase its benefits.
A. Category 1: EMT in Data Science Pipelines
Many aspects of data science and machine learning (ML) pipelines
benefit from the salient features of EAs for optimization.
Problems such as feature selection [46], hyper-parameter tuning
[47], neural architecture search [48], etc., involve non-differentiable,
multimodal objective functions and discrete search spaces
that call for gradient-free optimization. Population-based EAs
have even been considered as worthy rivals to, or in synergy
with, stochastic gradient descent for learning with differentiable
loss functions [49], [50]. Despite the advances, there remain
challenges in the efficient scaling of EAs to scenarios characterized
by big data (e.g., containing a large number of individual
data points), large-scale (high-dimensional) feature/parameter
spaces, or involving building sets of multiple learning algorithms
(e.g., ensemble learning). EMT provides different pathways
to sustain the computational tractability of EAs in such
data science settings.
EMT with auxiliary task generation: Approaches to augment
the training of ML models by turning the problem into
MTO-with artificially generated auxiliary tasks-were
IEEE Computational Intelligence Magazine - May 2022

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