Evolutionary Multitask Optimization in Real-World Applications: A Survey

Yue Wu, Hangqi Ding, Benhua Xiang, Jinlong Sheng, Wenping MaJournal of Artificial Intelligence and Technology 2023 · DOI

Abstract

Due to its good ability to solve problems, evolutionary multitask optimization (EMTO) algorithm has been widely studied recently. Evolutionary algorithm has the advantage of fast searching for the optimal solution, but it is easy to fall into local optimum and difficult to generalize. To solve these problems, it is an effective method to combine with multitask optimization algorithm. Through the implicit parallelism of tasks themselves and the knowledge transfer between tasks, more promising individuals can be generated in the evolution process, which can jump out of the local optimum. How to better combine the two has also been studied more and more. This paper will explore the existing evolutionary multitasking theory and improvement scheme in detail. Then it summarizes the application of evolutionary multitask optimization in different scenarios. Finally, according to the existing research, the future research trends and potential exploration directions are revealed.

Introduction

Traditional Evolutionary Algorithms (EAs) are renowned for their ability to navigate large search spaces to find optimal solutions. However, they frequently encounter two significant hurdles: the tendency to become trapped in local optima and a lack of generalization across related problems. In practical engineering, problems rarely exist in isolation; they often share underlying structures or features.

Evolutionary Multitask Optimization (EMTO) bridges the gap between theoretical EA research and complex, multi-faceted real-world problems. By treating multiple related problems as "tasks" within a single evolutionary framework, EMTO leverages the search experience gained in one task to assist others. This synergy allows the algorithm to effectively "map" the search space, helping individual tasks escape local traps through shared knowledge.

The EMTO Framework and Methodology

The core engine behind most EMTO research is the Multifactorial Optimization (MFO) paradigm. This approach allows a single population of agents to solve multiple optimization problems simultaneously by utilizing several key concepts:

  • Unified Search Space: All tasks are mapped into a single, high-dimensional space, enabling direct comparison and interaction between individuals assigned to different tasks.
  • Skill Factor: Each individual in the population is assigned a "skill factor" based on the specific task it performs most effectively.
  • Knowledge Transfer Mechanisms: Knowledge is shared through Assortative Mating, where individuals from different tasks can mate to pass on useful genetic traits, and Vertical Cultural Transmission, where offspring inherit the task expertise (skill factor) of their parents.

Advanced Iterations

Recent research has introduced more sophisticated versions of the original Multifactorial Evolutionary Algorithm (MFEA) to enhance performance and reliability:

  • MFEA-II: Incorporates online learning to track task similarity in real-time, preventing "negative transfer" where unrelated tasks might degrade each other's performance.
  • SREMTO: Dynamically adjusts transfer levels based on the evolving similarity of the search processes.
  • Explicit EMTO: Moves beyond random genetic crossover by building explicit mathematical mappings (such as sparse reconstruction) between tasks to transfer knowledge with higher precision.

Benchmarks and Real-World Applications

EMTO has been validated across a diverse range of benchmarks and practical scenarios, demonstrating its versatility:

  • Combinatorial Optimization: Success has been noted in classic problems like the Traveling Salesman Problem (TSP), Quadratic Assignment Problem (QAP), and Job Shop Scheduling (JSP).
  • Image and Signal Processing: EMTO is used for sparse unmixing in hyperspectral images and multimodal MRI scans for brain tumor segmentation.
  • Engineering and Robotics: In 3D point cloud registration, EMTO-based methods (like EMTR-SSC) show higher success rates and better noise robustness because auxiliary tasks help the main task avoid local optima.
  • Dynamic Scheduling: EMTO-based Genetic Programming (GP) demonstrates superior generalization across varying workshop scenarios compared to traditional single-task approaches.

Key Insights

One of the most profound insights in EMTO research is Implicit Parallelism. Rather than simply solving tasks concurrently, EMTO uses the process of solving one task to discover shortcuts for another, creating a more efficient overall search.

However, the field also highlights the "Negative Transfer" hurdle. Forcing information sharing between completely unrelated tasks can be counterproductive. Consequently, modern EMTO research focuses heavily on "task relationship learning" to ensure that knowledge transfer is beneficial. Furthermore, in multi-objective problems, EMTO can treat different objectives as separate tasks, maintaining population diversity and finding a more robust Pareto frontier.

Future Directions and Limitations

Despite its successes, several challenges remain for the future of EMTO:

  1. Theoretical Foundation: A more granular understanding of how and why knowledge transfers effectively is still needed to build a complete theoretical framework.
  2. Computational Efficiency: The overhead of calculating task similarities and mapping spaces can be significant. Future research must focus on developing "low-cost" transfer methods.
  3. Standardization: The lack of universal performance indices for multitask scenarios makes it difficult to compare different algorithms fairly.
  4. Broader Deployment: While established in engineering, EMTO has significant potential for expansion into fields like large-scale big data analytics and advanced medical diagnostics.

Summary updated

Abstract

Due to its good ability to solve problems, evolutionary multitask optimization (EMTO) algorithm has been widely studied recently. Evolutionary algorithm has the advantage of fast searching for the optimal solution, but it is easy to fall into local optimum and difficult to generalize. To solve these problems, it is an effective method to combine with multitask optimization algorithm. Through the implicit parallelism of tasks themselves and the knowledge transfer between tasks, more promising individuals can be generated in the evolution process, which can jump out of the local optimum. How to better combine the two has also been studied more and more. This paper will explore the existing evolutionary multitasking theory and improvement scheme in detail. Then it summarizes the application of evolutionary multitask optimization in different scenarios. Finally, according to the existing research, the future research trends and potential exploration directions are revealed.
BibTeX
@article{wu2023evolutionary,
  author = {Yue Wu and Hangqi Ding and Benhua Xiang and Jinlong Sheng and Wenping Ma},
  title = {{Evolutionary Multitask Optimization in Real-World Applications: A Survey}},
  journal = {Journal of Artificial Intelligence and Technology},
  year = {2023},
  doi = {10.37965/jait.2023.0149},
}
Figure