Transferable Adaptive Differential Evolution for Many-Task Optimization

Sheng-Hao Wu, Zhi-Hui Zhan, Kay Chen Tan, Jun ZhangIEEE Trans. Cybernetics 2023 · DOI

Introduction

Evolutionary Multitask Optimization (EMTO) aims to solve multiple optimization tasks simultaneously by leveraging shared knowledge. However, as the number of tasks scales to "many-task" optimization (MaTOP, typically 50+ tasks), traditional methods face significant challenges. Most existing approaches rely on Population Distribution-based Similarity Measurement (PDSM), which assumes tasks are similar only if their populations cluster in the same search space region. This fails to account for shift invariance—where tasks have identical landscapes but different global optima. Furthermore, traditional "solution transfer" often leads to negative transfer in complex many-task scenarios.

Shift invariance and the proposed SEET method compared to existing approaches.

To address these issues, the TRADE (Two-stage Transferable Adaptive Differential Evolution) algorithm is proposed. It shifts the focus from transferring individuals to transferring successful evolution "experiences"—specifically, algorithm parameters—while grouping tasks based on landscape similarity rather than spatial proximity.

The TRADE Framework

The TRADE algorithm operates in two distinct stages to manage the complexity of many-task environments effectively.

The two-stage framework of the TRADE algorithm.

Stage 1: Task Representation and Grouping

The first stage identifies relationships between tasks without prior knowledge of their structure:

  • Task Representation Strategy (TRS): Tasks are initially evolved independently. The "best-so-far" fitness history is recorded, serving as a unique signature of the task's landscape geometry.
  • Task Grouping Strategy (TGS): Using the Chinese Restaurant Process (CRP), TRADE automatically clusters tasks with similar fitness signatures. This non-parametric clustering allows the algorithm to discover the optimal number of groups dynamically, capturing shift-invariant similarities that spatial methods miss.

Stage 2: Successful Evolution Experience Transfer (SEET)

Once tasks are grouped, TRADE facilitates knowledge sharing through parameter transfer rather than solution swapping:

  • Evolution Quality Analysis (EQA): The algorithm ranks the performance of Differential Evolution (DE) parameters ($F$ and $Cr$) within each group.
  • Experience Transfer: "Successful" parameters from high-performing tasks are probabilistically transferred to struggling tasks within the same group. This adaptive strategy ensures that tasks benefit from the "how-to-solve" knowledge of their peers, reducing the risk of negative transfer.

Experimental Evaluation

The researchers evaluated TRADE against state-of-the-art EMTO algorithms (e.g., EMaTO, MaTDE) using the CEC19MaTOP and GECCO20MaTOP benchmarks, each featuring 50 tasks.

  • Benchmark Performance: TRADE significantly outperformed single-task and multi-task variants across nearly all instances. In heterogeneous benchmarks (GECCO20), TRADE successfully grouped tasks by their underlying function types (e.g., Ackley vs. Rastrigin), leading to superior convergence rates.
  • Real-World Application: In a 50-task robotic arm control problem, TRADE identified physical similarities in target distances from the base, grouping related tasks to accelerate the search for optimal joint angles.

Conclusion

TRADE introduces a robust paradigm for many-task optimization by prioritizing landscape-based similarity and strategy-level transfer. By identifying shift invariance and adaptively sharing successful evolution parameters, it overcomes the limitations of distribution-based similarity and the pitfalls of negative transfer. Future work will explore expanding similarity measures to include rotation invariance and refining task representation for noisier optimization environments.

Summary updated

BibTeX
@article{wu2023transferable,
  author = {Sheng-Hao Wu and Zhi-Hui Zhan and Kay Chen Tan and Jun Zhang},
  title = {{Transferable Adaptive Differential Evolution for Many-Task Optimization}},
  journal = {IEEE Trans. Cybernetics},
  year = {2023},
  doi = {10.1109/tcyb.2023.3234969},
}
Figure