Diffusion-Based Knowledge Transfer for Multitask Optimization

Area of Science:

• Artificial Intelligence
• Machine Learning
• Optimization

Background:

• Evolutionary multitask optimization (EMTO) seeks to optimize multiple tasks concurrently.
• Effective knowledge transfer (KT) is vital for EMTO success.
• Existing KT methods in EMTO offer limited knowledge transfer capabilities.

Purpose of the Study:

• To propose a novel diffusion model-based knowledge transfer (DMKT) method for EMTO.
• To address the limitations of superficial KT in traditional EMTO approaches.
• To deeply mine mapping relationships between tasks for improved knowledge transfer.

Main Methods:

• A diffusion model-based knowledge transfer (DMKT) method utilizing cold diffusion (CD) is proposed.
• Cyclic training of two diffusion models (DMs) for each source-target task pair.
• DMKT generates new solutions by learning directed cross-task mappings via degradation and restoration processes.

Main Results:

• The proposed diffusion-based KT multitask optimization (DKTMTO) algorithm demonstrated superior performance against state-of-the-art EMTO algorithms on the CEC2022 MTOP benchmark.
• DMKT integration improved the performance of other EMTO algorithms.
• DKTMTO showed applicability and scalability on real-world PKACPs and the WCCI2020 MaTOP benchmark.

Conclusions:

• DMKT offers a powerful approach for deep knowledge mining and transfer in EMTO.
• The DKTMTO algorithm represents a significant advancement in multitask optimization.
• The proposed method is effective and scalable for both benchmark and real-world applications.