referred to [33]. The key characteristic of these two processes is that they possess partially overlapping design spaces. Specifically, there exist three design parameters-the pressure and temperature of the epoxy resin when injected into the mold, and the temperature of the mold itself-that have similar physical effect on both LCM processes, hence leading to the scope of exploitable inter-task synergies. The RTM and I/C-LCM optimization problem instances were formulated as bi-objective minimization tasks. The first objective was to minimize mold filling time (which in turn increases process throughput), while the second was to minimize peak internal fluid and fiber compaction force (which in turn reduces setup and running cost of peripheral equipment). VRP1 3 2 4 14 5 1 16 17 12 13 11 15 14 (a) 8 7 10 6 7 9 11 9 6 8 12 10 5 13 1 15 4 3 2 For a set of candidate design parameters, the objective function values for either task were evaluated using a dedicated finite element numerical simulation engine. The outputs of the multitasking MO-MFEA and the single-task NSGA-II are compared in Fig. 6 in terms of the normalized hypervolume metric. The convergence trends achieved by MO-MFEA on both tasks were found to surpass those achieved by NSGA-II. Taking RTM as an example (see left panel of Fig. 6), the MO-MFEA took only about 1000 evaluations to reach the same hypervolume score reached by NSGAII at the end of 2000 evaluations. This represents a ~50% saving in cost, which for expensive simulation-based optimization problems (ubiquitous in scientific and engineering applications) VRP2 Solution 7 Representation Learning Scaling 5 4 Rotation 4 16 Translation 1 3 14 2 (b) (c) FIGURE 7 (a) VRP1 and VRP2 possess seemingly dissimilar node distribution and labels; (b) solution representation learning is undertaken to isometrically transform the node distribution of VRP2 to match VRP1; (c) the similarity of the two VRPs is unveiled after the transformation [21]. 15 10 10 15 13 11 12 Depot Customer of VRP1 Customer of VRP2 6 17 9 8 MAY 2022 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 59