Block-Level Knowledge Transfer for Evolutionary Multitask Optimization

Yi Jiang, Zhi‐Hui Zhan, Kay Chen Tan, Jun Zhang

Abstract

Evolutionary multitask optimization (EMTO) is an emerging field focusing on solving multiple optimization tasks concurrently by sharing common knowledge. However, existing knowledge transfer mechanisms are often limited to aligned dimensions between tasks and frequently overlook correlations within the same task. This article proposes the Block-Level Knowledge Transfer (BLKT) framework to address these limitations. BLKT partitions individuals into blocks and clusters similar blocks—regardless of their task or dimension origin—to facilitate more rational knowledge transfer. Integrated with Differential Evolution (BLKT-DE), the framework demonstrates superior performance across CEC benchmarks and real-world kinematic arm control problems.

The Alignment Trap in EMTO

The primary goal of EMTO is to leverage insights from one task to accelerate the optimization of others. Despite its potential, traditional EMTO algorithms often fall into two "alignment traps":

  1. The Index Trap: Knowledge transfer is typically restricted to dimensions with identical indices. This assumes that dimension $i$ in Task A is inherently related to dimension $i$ in Task B, which is often not the case in complex, real-world scenarios.
  2. Intratask Neglect: Most frameworks focus exclusively on inter-task transfer, ignoring valuable correlations between different dimensions within a single task.

These constraints can lead to "negative transfer," where the exchange of irrelevant information hinders the optimization process.

Comparison of traditional index-based transfer (a) vs. the proposed block-level transfer (c, d) which handles unaligned dimensions and intratask relationships.

The BLKT Framework

To overcome these hurdles, the BLKT framework introduces a more flexible, block-based approach to knowledge transfer. Instead of transferring information between whole individuals or fixed indices, BLKT operates on segments of the solution space.

The BLKT-DE algorithm follows a cyclic five-step process:

  1. Divide: Each individual is partitioned into multiple blocks of a specified length $B$.
  2. Get Block Population: Blocks from all tasks are aggregated into a single global block population.
  3. K-means Clustering: The algorithm groups similar blocks into $K$ clusters. This step is crucial as it identifies related dimensions dynamically, whether they belong to the same task or different ones.
  4. Mutation and Crossover: Evolutionary operations occur within these clusters. Since the blocks in a cluster are mathematically similar, the resulting knowledge transfer is highly relevant and effective.
  5. Reconstruct: Evolved blocks are reassembled into their original positions to form the next generation of individuals.
The five-step BLKT process: dividing individuals into blocks, clustering similar blocks, evolving within clusters, and reconstructing individuals.

To maintain efficiency, the authors implemented a Feedback-based Adaptive Strategy (FAS) to dynamically adjust the hyperparameters $B$ (block size) and $K$ (number of clusters) based on the algorithm's performance in each generation.

Experimental Results

The BLKT-DE framework was rigorously tested against state-of-the-art algorithms on several benchmarks:

  • CEC17M & CEC22M: BLKT-DE consistently outperformed competitors like MFEA-AKT and ASCMFDE.
  • cMTOP: On this "compositive" suite designed specifically to test unaligned dimensions, BLKT-DE showed significant robustness, winning 19 out of 20 task comparisons.
  • Real-World Scaling: In planar kinematic arm control tasks, the framework maintained peak performance as the number of tasks scaled from 2 to 100.
  • Single-Task Optimization: Interestingly, BLKT-DE also proved competitive in single-task scenarios (CEC17S/22S), suggesting that treating internal problem components as "tasks" can accelerate convergence.

Conclusion

The BLKT framework represents a significant shift in evolutionary multitask optimization by prioritizing functional similarity over arbitrary indexing. By enabling both inter-task and intra-task transfer at the block level, it provides a more robust and flexible mechanism for knowledge sharing. Future research may explore non-consecutive block structures and the integration of BLKT with other metaheuristic algorithms to further broaden its applicability.

Summary updated

Abstract

Evolutionary multitask optimization is an emerging research topic that aims to solve multiple tasks simultaneously. A general challenge in solving multitask optimization problems (MTOPs) is how to effectively transfer common knowledge between/among tasks. However, knowledge transfer in existing algorithms generally has two limitations. First, knowledge is only transferred between the aligned dimensions of different tasks rather than between similar or related dimensions. Second, the knowledge transfer among the related dimensions belonging to the same task is ignored. To overcome these two limitations, this article proposes an interesting and efficient idea that divides individuals into multiple blocks and transfers knowledge at the block-level, called the block-level knowledge transfer (BLKT) framework. BLKT divides the individuals of all the tasks into multiple blocks to obtain a block-based population, where each block corresponds to several consecutive dimensions. Similar blocks coming from either the same task or different tasks are grouped into the same cluster to evolve. In this way, BLKT enables the transfer of knowledge between similar dimensions that are originally either aligned or unaligned or belong to either the same task or different tasks, which is more rational. Extensive experiments conducted on CEC17 and CEC22 MTOP benchmarks, a new and more challenging compositive MTOP test suite, and real-world MTOPs all show that the performance of BLKT-based differential evolution (BLKT-DE) is superior to the compared state-of-the-art algorithms. In addition, another interesting finding is that the BLKT-DE is also promising in solving single-task global optimization problems, achieving competitive performance with some state-of-the-art algorithms.
BibTeX
@article{jiang2023block,
  author = {Yi Jiang and Zhi‐Hui Zhan and Kay Chen Tan and Jun Zhang},
  title = {{Block-Level Knowledge Transfer for Evolutionary Multitask Optimization}},
  journal = {IEEE Transactions on Cybernetics},
  year = {2023},
  doi = {10.1109/tcyb.2023.3273625},
}
Figure