IEEE Computational Intelligence Magazine - May 2022 - 60

... there are still challenges in reliably implementing
EMT for combinatorial optimization tasks .... A key
issue is that of solution representational mismatch,
which can lead to negative transfer.
translates to substantial reduction in design time and the consumption
of valuable physical resources.
E. Category 5: EMT in Manufacturing,
Operations Research
The grand vision of smart manufacturing involves integration of
three levels of manufacturing systems, namely, the shop floor,
enterprise, and supply chain, into automated and flexible networks
that allow for seamless data collection (via distributed sensors),
data exchange, analysis, and decision-making [98]. These
may be supported by a nerve center or manufacturing control tower,
where real-time data is collected across all system levels to offer
centralized processing capacity and end-to-end visibility. It is in
enabling effective functioning of such control towers that we
foresee EMT to thrive, leveraging the scope of seamless data
exchanges to deliver fast and optimal (or near-optimal) operational
decisions [99].
Targeting energy efficient data collection and transmission
to the base location (e.g., the nerve center), [100] demonstrated
the utility of EMT for optimizing the topology of wireless
sensor networks. The optimization of both single-hop and
multi-hop network types were combined in MTO to help
with consideration of both deployment options. It was shown
using a variant of the MFEA with random-key encoding that
the exchange of useful information derived from solving both
tasks could in fact lead to better overall results than the baseline
single-task method. In [101], the follow-on problem of
charging the wireless sensors was also undertaken using a multitask
approach. Multiple mobile chargers were simultaneously
considered, with the charging schedule for each forming a task
in MTO.
Returning to manufacturing operations, there exists a sizeable
amount of research on applying EMT algorithms to NPhard
problems at the shop floor (e.g., for job shop scheduling
[102], [103]) or at the logistics and supply chain levels (e.g., for
vehicle routing applications [104], [105] and its extension to
pollution-routing [106]). For last-mile logistics in particular,
centralized cloud-based EMT was envisioned in [8], [107] to
take advantage of similarities in the graph structures of vehicle
routing problem (VRP) instances toward rapid optimization.
The application of EMT to other forms of graph-based optimization
tasks with potential use in manufacturing have also
been explored in [108], [109].
Despite some success, there are still challenges in reliably
implementing EMT for combinatorial optimization tasks ubiquitous
in manufacturing and operations research. A key issue is
that of solution representational mismatch, which can lead to
60 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2022
negative transfer [110]. For instance, consider
unifying two VRPs in EMT that are defined
using different node labels/indices even
though their underlying customer distributions
are similar. Due to the resultant label
mismatch, genetic transfers under standard
permutation-based solution representations
would lead to suboptimal (or even confounding)
exchange of routes or subroutes between tasks.
Two recent research avenues hold promise in overcoming
the aforementioned challenge. The first entails departure from
the usual direct transfer of solution prototypes in EMT. Instead,
the transfer of higher-order solution construction heuristics that are
agnostic to low-level solution representations is proposed (as a
form of multitask hyper-heuristic); both heuristic selection
[18] and generative approaches [111] have been put forward,
showing greater robustness to representational mismatches in
EMT. The second research avenue deals with learning solution
representations, transforming problem instances in a manner
that minimizes inter-task representational mismatch. An illustration
of this idea is depicted in Fig. 7, where two VRP
instances (VRP1
and VRP )2
with seemingly dissimilar customer
distributions and node labelling are examined. However,
through an isometric transformation (comprising rotation
and translation operations) of the nodes in VRP2
(which preserves
shortest routes), a new representation scheme that better
aligns both tasks is obtained [21].
1) Case study in last-mile logistics planning [112]
Following on from the discussions above, a case study on realworld
package delivery problem (PDP) instances [112] from a
courier company in Beijing, China, is presented. The PDP is a
variant of the NP-hard VRP, where the objective function pertains
to minimizing total routing costs in servicing a set of geographically
distributed customers (as illustrated in Fig. 7) with a
fleet of capacity constrained vehicles located at a single or multiple
depots. The results presented hereafter are for an explicit
EMT combinatorial optimization algorithm (EEMTA for
short) whose uniqueness lies in incorporating solution representation
learning via sparse matrix transformations to facilitate
the transfer of useful information across tasks. The reader is
referred to [112] for full details of the EEMTA and the algorithmic
settings used in the experimental study.
The experiments were conducted on four PDP requests
that were paired to form two examples of MTO. The pairing
was done based on customer distributions, with the resulting
MTO formulations referred to as {PDP1, PDP2} and {PDP3,
PDP4}, respectively. The convergence trends achieved by the
EEMTA and the baseline single-task EA (hybridized with
local search heuristics) are presented in Fig. 8. As can be seen
in the figure, the EEMTA was able to achieve some extent of
performance speed up across all four tasks. Multitasking provided
an impetus to the overall search, whilst strongly boosting
outcomes of the initial stages of evolution on PDP2 and
PDP4 in particular.
IEEE Computational Intelligence Magazine - May 2022

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