Half a Dozen Real-World Applications of Evolutionary Multitasking, and More

I. INTRODUCTION

• EMT jointly tackles distinct, possibly related tasks, enabling inter-task transfer and adaptive reuse of information that conventional single-task EAs leave inaccessible.The paper notes that real-world problems seldom exist in isolation and that existing EMT methods raise questions about what, how, and when to transfer.Source: §I · p. 1
• EMT research includes single-population multifactorial, multi-population, and co-evolutionary approaches, with choices about what, how, and when to transfer.The paper directs readers to other reviews for a more comprehensive methodological analysis.Source: §I · pp. 1–2
• The paper reviews real-world EMT explorations in six broad application categories, with representative experimental studies, while noting that the examples are not comprehensive.The examples include embodied intelligence, unmanned-vehicle path planning, and last-mile logistics optimization.Source: §I · p. 2
• It also offers recipes for recasting cross-disciplinary problem formulations as multi-X evolutionary multitasking problems.These formulations are presented as ways to leverage population implicit parallelism and motivate future algorithms for real-world deployment.Source: §I · p. 2

A. The Multitask Optimization Problem

• Multitask optimization jointly solves K problem instances, typically encoding their candidate solutions in a unified space to support transfer across tasks.Task-specific mappings translate solutions into and out of that shared space; learned representations can support transfers between heterogeneous search spaces.Source: A. The Multitask Optimization Problem · p. 2

B. A Probabilistic Formulation of EMT

• A task’s fitness can be expressed as expected fitness under a probability distribution, which represents the evolving population’s search distribution.A Dirac distribution centered on the true optimum optimizes this expected-fitness formulation.Source: B. A Probabilistic Formulation of EMT · p. 2
• EMT generalizes the formulation across K tasks using a mixture of task-specific models, with weights controlling how much each model contributes.Increasing a source-to-target weight intensifies cross-sampling when its solutions perform well for the target; ineffective sources are gradually neutralized.Source: B. A Probabilistic Formulation of EMT · pp. 2–3

C. An Overview of EMT Methodologies

• Implicit EMT transfers information when candidate solutions from different tasks cross over in a unified space, exchanging skills encoded in their genetic material.Its success depends on whether the crossover operator can reveal and exploit relationships between the tasks’ objective landscapes.Source: C. An Overview of EMT Methodologies · p. 3
• Explicit EMT evolves separate task-specific populations and triggers knowledge exchange between them under a predefined condition, such as a transfer interval.Heterogeneous search spaces require mappings between populations; existing approaches often use linear mappings, while nonlinear maps are identified as a future direction.Source: C. An Overview of EMT Methodologies · p. 3
• Implicit transfer is relatively simple and scales well for large K, but unifying heterogeneous tasks is difficult; explicit transfer permits specialized operators but can require O(K^2) mappings.The paper recommends choosing between the approaches based on application attributes rather than treating either as universally preferable.Source: C. An Overview of EMT Methodologies · p. 3

A. Category 1: EMT in Data Science Pipelines

• Data-science applications include feature selection, hyperparameter tuning, and neural architecture search, where EMT is discussed as a pathway to address EA scaling challenges.The challenges include big data, high-dimensional search spaces, and sets of learning algorithms such as ensembles.Source: A. Category 1: EMT in Data Science… · p. 3
• Auxiliary small-data tasks can accelerate evolutionary configuration on a large dataset; some wrapper-based feature-selection cases achieved speedups over 40%.Related approaches also use different losses or network topologies, or low-dimensional tasks to lessen stagnation in high-dimensional feature spaces.Source: A. Category 1: EMT in Data Science… · pp. 3–4
• Multifactorial genetic programming jointly evolves an ensemble of decision trees, reusing common subtrees to reduce the cost of repeated independent runs.Other EMT feature-selection work jointly identifies feature subspaces for ensemble learners, which often outperformed independently optimized subspaces.Source: A. Category 1: EMT in Data Science… · p. 4
• For hyperspectral images, EMT uses homogeneous regions as sparse-regression tasks to exploit shared sparsity patterns, with reported faster convergence to near-optimal solutions.A multi-fidelity procedure also used a surrogate to estimate the gap between low- and high-fidelity evaluations, improving accuracy and algorithmic efficiency.Source: A. Category 1: EMT in Data Science… · p. 4
• Cloud-based model configuration is proposed as a setting for transferring knowledge across related tasks with different training data or device requirements.A figure depicts users uploading raw data to a service that returns optimized predictive models; early plausibility studies used multitask Bayesian optimization.Source: A. Category 1: EMT in Data Science… · p. 4, Figure 1 · p. 4

• Case study in symbolic regression modeling [68]

• MFGP outperformed SL-GEP in all three symbolic-regression settings, with the best CO2 and DRP results paired with their corresponding simplified variants [68].The simplified tasks were generated by subsampling the original time series.Source: • Case study in symbolic regression… · p. 5

B. Category 2: EMT in Evolving Embodied Intelligence

• An EMT algorithm significantly outperformed a strong single-task baseline while evolving walking controllers for 2,000 randomly parameterized six-legged robot morphologies.Each task corresponded to a different set of leg lengths.Source: B. Category 2: EMT in Evolving… · p. 5
• For variable-morphology planar-arm articulation, adaptive EMT achieved faster convergence and better average objective values than a single-task EA.The tasks varied arm-link lengths and sought joint angles that minimized the distance to a target.Source: B. Category 2: EMT in Evolving… · p. 5
• Across up to 50 Meta-World tasks, crossover-based skill exchange among synergistic tasks yielded higher success rates and lower computational cost than single-task soft actor critic [69].The tasks evolved policy parameters for a robot arm interacting with objects that differed in shape, joints, and connectivity.Source: B. Category 2: EMT in Evolving… · p. 6

• Case study in neuroevolution of robot controllers [26]

• The pole-balancing study varied the shorter pole among 0.60 m, 0.65 m, and 0.70 m, using a shared two-layer controller with ten hidden neurons across tasks [26].The long pole was fixed at 1 m, and the study considered four task combinations.Source: • Case study in neuroevolution of… · p. 6

C. Category 3: EMT in Unmanned Systems Planning

• UAV flight paths can share reusable segments because similar surroundings, such as buildings, shape their routes.The illustration identifies shared segments across different paths as potential transfer building blocks.Source: Figure 3 · p. 6, C. Category 3: EMT in Unmanned Systems… · p. 6

• Case study in neuroevolution of robot controllers [26]

• MFEA-II improved success rates over CEA on all three pole-balancing tasks, reaching 47% on T1, 37% on T2, and 27% on T3.CEA achieved 27% on T1 and failed on T2 and T3; the two algorithms used the same operators.Source: • Case study in neuroevolution of… · p. 6

C. Category 3: EMT in Unmanned Systems Planning

• Different UAV missions in had a shared optimal deviation magnitude from straight-line paths, which MFEA exploited to evolve efficient flight paths.The missions involved either two UAVs passing through barrier openings or four UAVs routing around a geofence.Source: C. Category 3: EMT in Unmanned Systems… · p. 6
• Both MO-MFEA-II settings achieved better hypervolume scores faster than the single-task EA, with greater speedup when two auxiliary tasks were used [84].The auxiliary tasks had lower fidelity than the high-fidelity target T1.Source: Figure 4 · p. 7
• EMT supported efficient obstacle-avoiding trajectories for mobile agents, while path-information transfer across three terrains produced shorter, safer path sets faster than a single-task counterpart.The terrains had distinct topographic maps but similar obstacle distributions.Source: C. Category 3: EMT in Unmanned Systems… · p. 7

• Case study in safe multi-UAV path planning [84]

• MO-MFEA-II with {T1, T2} and {T1, T2, T3} achieved better hypervolume scores faster than a single-task multi-objective EA, with the greater speedup using two auxiliary tasks.The five-UAV case minimizes travel distance and unsafe-event probability while satisfying separation, altitude, and geofence constraints; lower-resolution meshes provide cheaper auxiliary tasks.Source: • Case study in safe multi-UAV path… · p. 7

D. Category 4: EMT in Complex Design

• Across scientific and engineering design, multitask optimization can seed searches with information from related tasks, an approach suited to costly simulations and laboratory evaluations.The paper notes potential value during conceptualization, when multiple concepts with latent synergies are assessed together.Source: D. Category 4: EMT in Complex Design · p. 7
• For multitask aerodynamic car-shape optimization, a 3D point-cloud autoencoder provided a shared latent space for transferring solution building-blocks, yielding up to 38.95% lower drag than a single-task baseline under the same computational budget.Source: D. Category 4: EMT in Complex Design · pp. 7–8
• For continuous annealing production processes, AdaMOMFDE significantly outperformed single-task NSGA-II on convergence trends while meeting design specifications and potentially boosting productivity.The study jointly solved up to eight tasks, each with hardness, energy-consumption, and production-capacity objectives.Source: D. Category 4: EMT in Complex Design · p. 8

• Case study in simulation-based process design [33]

• MO-MFEA surpassed NSGA-II on both RTM and I/C-LCM tasks in normalized hypervolume convergence; on RTM, it reached NSGA-II’s end-of-run score in about 1,000 rather than 2,000 evaluations.The tasks optimize mould filling time and peak internal fluid and fibre compaction force using finite-element simulations; the RTM comparison represents about 50% evaluation-cost savings.Source: • Case study in simulation-based… · p. 8, Figure 6 · p. 8

E. Category 5: EMT in Manufacturing, Operations Research

• EMT is proposed as a way to support manufacturing control towers by leveraging data exchange for fast, optimal or near-optimal operational decisions.The envisioned setting integrates shop-floor, enterprise, and supply-chain systems.Source: E. Category 5: EMT in Manufacturing,… · p. 9
• A random-key MFEA variant improved on a single-task baseline by jointly optimizing single-hop and multi-hop wireless sensor network topologies; multitask formulations also addressed mobile-charger schedules.The network-optimization goal was energy-efficient data collection and transmission to a base location.Source: E. Category 5: EMT in Manufacturing,… · p. 9
• EMT applications in manufacturing operations include shop-floor job-shop scheduling, vehicle routing, and pollution routing, with cloud-based EMT envisioned for last-mile logistics.These approaches exploit similarities among graph structures or tasks to support rapid optimization.Source: E. Category 5: EMT in Manufacturing,… · p. 9
• Solution representational mismatch can lead to negative transfer and remains a challenge for reliably applying EMT to combinatorial optimization in manufacturing and operations research.When similar VRPs use different node labels, standard permutation-based transfers can exchange routes or subroutes suboptimally or confoundingly.Source: E. Category 5: EMT in Manufacturing,… · p. 9
• Higher-order construction heuristics and learned solution representations are two proposed ways to improve robustness to representational mismatch.The cited representation-learning example aligns VRP node distributions through rotation and translation while preserving shortest routes.Source: E. Category 5: EMT in Manufacturing,… · p. 9

• Case study in last-mile logistics planning [112]

• EEMTA, which uses sparse-matrix solution representation learning to transfer information across tasks, achieved some performance speedup across four Beijing package-delivery requests paired into two multitask formulations.The largest early-stage boosts were reported on PDP2 and PDP4; Fig. 8 compares convergence with a single-task EA hybridized with local search.Source: • Case study in last-mile logistics… · pp. 9–10, Figure 8 · p. 10

F. Category 6: EMT in Software and Services Computing

• For web-service composition, three MFEA variants required significantly less execution time than a state-of-the-art single-task EA while achieving competitive and sometimes better semantic-matchmaking and quality-of-service results.The multitask formulation jointly handles similar composition requests, with solutions encoding service couplings into execution workflows.Source: F. Category 6: EMT in Software and… · p. 10
• In software branch testing, MTEC achieved competitive or superior coverage on most programs compared with independently searching each branch under a uniform computational budget.The study averaged coverage over 20 runs on 10 C numerical-calculus functions; infeasible branches can make optimal coverage less than 100%.Source: F. Category 6: EMT in Software and… · p. 10, Table III · p. 10
• Similar abstract syntax trees in different programs offer common code subtrees that an imagined automated program generator could transfer and reuse through EMT.Source: Figure 9 · p. 11

IV. FORGING A FUTURE FOR EMT IN MULTI-X PROBLEMS

• The paper presents cross-disciplinary problem formulations as future EMT recipes intended to encourage research on population-based search with implicit and explicit multitasking strategies.Source: §IV · p. 11

A. EMT in Multi-Objective, Multi-Constrained Problems

• For many-objective problems, auxiliary multi-objective tasks can be generated so that their Pareto-front points also belong to the target problem’s Pareto front; a cited benchmark study reports outperforming state-of-the-art baselines under limited budget.For multi-constrained problems, the paper suggests dropping constraints in auxiliary tasks if the unknown active constraints are preserved.Source: A. EMT in Multi-Objective,… · p. 11

B. EMT in Multi-Fidelity Optimization

• Multi-fidelity EMT jointly evolves cheap, less accurate lower-fidelity tasks with an accurate, expensive target to reduce high-fidelity analysis effort.Examples cited include safe UAV path planning, beneficiation-process optimization, and multitasking across local and global surrogate models.Source: B. EMT in Multi-Fidelity Optimization · p. 11

C. EMT in Multi-Level Optimization

• In black-box bilevel optimization, each upper-level candidate can require repeated optimization of a lower-level problem, creating computational bottlenecks.EMT treats lower-level instances associated with upper-level candidates as tasks in a single multitasking setting.Source: C. EMT in Multi-Level Optimization · p. 12
• Approximately 65% computation time was saved over a standard evolutionary bilevel algorithm for lightweight-composite manufacturing, while surrogate-assisted MFEA vastly outperformed all baselines on robust airfoil design,.The airfoil problem used an expensive minimax formulation.Source: C. EMT in Multi-Level Optimization · p. 12

D. EMT in Multi-Scenario Optimization

• Fusing objectives across scenarios into one large multiobjective problem can impair tractability and return solutions outside individual scenarios’ Pareto sets [129].The proposed alternative treats each scenario as a separate task with post-hoc coordination.Source: D. EMT in Multi-Scenario Optimization · p. 12
• For intra-hour optimal power flow under rapid load variations, MFEA generated optimized scenario solutions proactively for use as corrective power-generation setpoints, improving operational economy.Source: D. EMT in Multi-Scenario Optimization · p. 12

• The review groups application-oriented studies into six broad categories and cites further work including UAV-swarm collision avoidance that outperformed state-of-the-art models.Source: §V · p. 12
• EMT’s transfer of multi-X optimization formulations into multitasking offers practitioners tools beyond jointly solving distinct problem instances.Source: §V · p. 12
• Reliable deployment requires consistent gains over state-of-the-art single-task methods, scalability to many tasks, and robustness to unrelated tasks to minimize performance degradation.The paper calls for rigorous comparisons across related, unrelated, and mixed task settings.Source: §V · pp. 12–13