Multitask Optimization in Evolutionary Computation

Multitask optimisation within evolutionary computation refers to the concurrent resolution of several optimisation tasks by a single algorithmic framework, leveraging shared information to accelerate convergence and enhance solution quality. This paradigm draws inspiration from human multitasking and biological evolution, where genetic material flows across related yet distinct challenges. Core to the approach is the notion of implicit or explicit knowledge transfer: valuable traits discovered in one task can inform the search in another, provided that underlying similarities are identified and exploited. Algorithms typically employ unified or partially unified representations, enabling cross-task recombination or migration of candidate solutions. Key benefits include reduced computational cost relative to separate searches, improved robustness against local optima through diversified genetic pools, and adaptability to problems with heterogeneous characteristics. Applications span engineering design, power distribution, machine learning hyperparameter tuning and complex scheduling. Despite promising results, challenges remain in measuring intertask similarity, preventing negative transfer when tasks diverge, and designing scalable strategies for many-task scenarios.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent advances have refined knowledge‐transfer mechanisms and adaptive strategies. A novel transferable adaptive differential evolution algorithm addresses many‐task optimisation by identifying shift invariance among tasks, grouping similar problems and selectively sharing successful evolution parameters to improve convergence on up to hundreds of tasks. Another line of work introduces orthogonal transfer via cross‐task mapping, enabling high‐quality transfer across tasks of differing dimensionalities; this mapping projects elite solutions into new search spaces, then applies an orthogonal recombination to ensure only beneficial dimensions are exchanged. A meta‐knowledge transfer differential evolution method further generalises transfer by evolving “knowledge of knowledge” rather than task‐specific traits, employing multiple populations in a unified search space alongside elite solution exchange to foster positive transfer across heterogeneous tasks. Collectively, these studies demonstrate enhanced performance on benchmark suites and real‐world applications by dynamically adapting transfer strategies to task characteristics.

Multitask Optimization in Evolutionary Computation publication trend

The graph below shows the total number of articles in multitask optimization in evolutionary computation across all publications each year (not limited to Nature Index journals).

Technical terms

Multitask optimisation: Simultaneous optimisation of multiple tasks within one framework to exploit shared information and improve overall efficiency.

Evolutionary computation: A family of population-based, nature-inspired algorithms that use mechanisms such as selection, recombination and mutation to evolve solutions to optimisation problems.

Knowledge transfer: The process of sharing algorithmic insights or solution components across different tasks to accelerate convergence and avoid redundant computation.

Multifactorial evolutionary algorithm (MFEA): A baseline evolutionary multitasking approach that assigns each individual a skill factor and applies cross-task crossover under a unified representation.

Cross-task mapping: A strategy for projecting solutions from one task’s search space into another, accommodating differences in dimensionality or variable domains.

Many-task optimisation (MaTOP): Extension of multitask optimisation to scenarios involving a large number of tasks, requiring scalable grouping and transfer mechanisms.

References

1. A Meta-Knowledge Transfer-Based Differential Evolution for Multitask Optimization. IEEE Transactions on Evolutionary Computation (2021).
2. Orthogonal Transfer for Multitask Optimization. IEEE Transactions on Evolutionary Computation (2022).
3. Transferable Adaptive Differential Evolution for Many-Task Optimization. IEEE Transactions on Cybernetics (2023).