Enhancing Generalization and Scalability for Multi-Objective Optimization with Population Pre-Training

Area of Science:

• Computational intelligence and machine learning
• Optimization algorithms and evolutionary computation

Background:

• Multi-objective optimization problems (MOPs) involve optimizing conflicting objectives, often with complex characteristics like high dimensionality and expensive evaluations.
• Existing population-based evolutionary computation methods for MOPs often lack generalizability across different problem types.

Purpose of the Study:

• To propose a Population Pre-trained Model (PPM) inspired by machine learning pre-training to efficiently solve complex MOPs within a unified framework.
• To address challenges in handling diverse decision spaces and capturing objective-decision space interdependencies during evolution.

Main Methods:

• Developed a population transformer architecture to embed decision spaces of varying scales into a common latent space for knowledge transfer.
• Integrated objective-space features through objective fusion to improve population prediction accuracy for complex MOPs.
• Leveraged historical optimization knowledge for pre-training the model.

Main Results:

• Achieved robust generalization to downstream optimization tasks with up to 5,000 dimensions, significantly exceeding prior work.
• Demonstrated consistent superiority over state-of-the-art algorithms on standardized benchmarks and real-world applications.
• Improved the performance and generalization capabilities of evolutionary computation for solving MOPs.

Conclusions:

• The proposed Population Pre-trained Model (PPM) offers a generalizable and efficient approach to tackling complex multi-objective optimization problems.
• PPM enhances evolutionary computation by enabling effective knowledge transfer and improving performance across diverse optimization tasks.