Diffusion-Based Knowledge Transfer for Multitask Optimization.

Abstract

Evolutionary multitask optimization (EMTO) aims to optimize multiple tasks simultaneously by transferring the related knowledge between tasks. Therefore, knowledge transfer (KT) between different tasks is crucial for facilitating the optimization of tasks. The traditional KT methods in EMTO achieve superficial KT through individual transfers, limiting their ability to transfer high-quality knowledge from other tasks. Therefore, in this article, a diffusion model-based KT (DMKT) method built specifically on cold diffusion (CD) is proposed to deeply mine the mapping relationships between different tasks and obtain transfer models for mapping individuals across tasks. Since CD supports arbitrary degradation operators, the interpolation between fitness-paired source and target task individuals can be defined as a task-oriented gradual degradation process, enabling the corresponding restoration process to learn directed cross-task mappings. In particular, we first construct two cyclic training diffusion models (DMs) for each source-target task pair. Subsequently, the trained DM can be used to generate new promising solutions for achieving efficient KT. The experimental results on the CEC2022 multitask optimization problem (MTOP) benchmark demonstrate that the proposed diffusion-based KT multitask optimization (DKTMTO) algorithm outperforms other state-of-the-art EMTO algorithms. Moreover, DMKT can be integrated into other EMTO algorithms to further improve their performance. Finally, DKTMTO is applied to real-world multitask planar kinematic arm control problems (PKACPs) and the WCCI2020 many-task optimization problem (MaTOP) benchmark, demonstrating its applicability and scalability.