IEEE Computational Intelligence Magazine - May 2022 - 58

differential evolution (AdaMOMFDE) algorithm was proposed for
optimizing continuous annealing production processes under different
environmental conditions. A set of environmental parameters
defined a certain steel strip production task, with multiple
parameter sets forming multiple problem instances in MTO. Each
task possessed three objectives, that of achieving prescribed strip
hardness specifications, minimization of energy consumption, and
maximization of production capacity. Experiments simultaneously
solving up to eight tasks were carried out in [96]. The results
demonstrated that the AdaMOMFDE algorithm could significantly
outperform the single-task NSGA-II (as quantified by convergence
trends of the inverted generational distance metric),
hence meeting design specifications while potentially boosting
productivity in the iron and steel industry.
In addition to the focused application areas above, MTO
lends a general framework for handling expensive design
optimizations by jointly incorporating tasks of multiple levels
of fidelity. The real-world case study in the previous subsection
was a case in point, albeit belonging to a different category.
Other related studies have also appeared in the literature [97], a
more extended discussion on which shall be presented in Section
IV-B of this paper.
1) Case study in simulation-based process design [33]
This study showcases an example where EMT was applied to
jointly optimize two types of liquid composite molding
(LCM) processes for producing the same lightweight composite
part [33]. The part under consideration was a glass-fiberreinforced
epoxy composite disk, while the two LCM
processes were resin transfer molding (RTM) and injection/
compression LCM (I/C-LCM). The process details are not
reproduced herein for the sake of brevity; interested readers are
Knowledge
Transfer
Unified
Representation
Reconstruction
FIGURE 5 In many applications of EMT for engineering design, the lack of clear semantic overlap between design parameters could lead to difficulties
in the construction of the unified search space
X . One example is in the definition of the unified space of diverse car shapes/geometries
for aerodynamic design, which was addressed in [29] using a 3D point cloud autoencoder. Once trained, inter-task knowledge transfers take
place in the latent space of the autoencoder.
RTM
I/C-LCM
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75
500
1,000
Evaluations
(a)
MO-MFEA
NSGA-II
1,500
2,000
0.4
0.45
0.5
0.55
0.6
0.65
0.7
Evaluations
(b)
FIGURE 6 (a) Hypervolume convergence trends of MO-MFEA and NSGA-II on the RTM process optimization task; (b) hypervolume convergence
trends of MO-MFEA and NSGA-II on the I/C-LCM process optimization task. These plots have been obtained from the real-world study in [33].
MO-MFEA
NSGA-II
0 200 400 600 800 1,000 1,200 1,400
58 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2022
Performance (Hypervolume)
Performance (Hypervolume)

IEEE Computational Intelligence Magazine - May 2022

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