Learning Evolution via Optimization Knowledge Adaptation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Evolutionary Computation

Background:

• Evolutionary algorithms (EAs) store optimization knowledge in historical data.
• Current methods isolate knowledge transfer and online adaptation, limiting effectiveness.
• Existing approaches struggle with complete knowledge utilization and operator-specific adaptations.

Purpose of the Study:

• To develop a unified framework for evolutionary algorithms that simultaneously enables knowledge transfer and online adaptation.
• To introduce the Optimization Knowledge Adaptation Evolutionary Model (OKAEM) for adaptive parameter updating.
• To leverage optimization knowledge for improved EA performance.

Main Methods:

• Introduced the Optimization Knowledge Adaptation Evolutionary Model (OKAEM), a learnable evolutionary framework.
• Parameterized evolutionary operators using attention mechanisms for adaptive updates.
• Implemented a two-phase approach: pre-training for knowledge transfer and adaptive optimization for real-time tuning.

Main Results:

• OKAEM significantly outperformed state-of-the-art sequential transfer methods in 12 transfer scenarios via pre-training.
• OKAEM surpassed advanced learnable EAs in prior-free settings using its self-tuning mechanism.
• Demonstrated practical utility in prompt tuning for vision-language models.

Conclusions:

• OKAEM effectively integrates and utilizes optimization knowledge for both transfer and adaptation.
• The model's learnable components and adaptive mechanisms are crucial for its superior performance.
• OKAEM autonomously discovers interpretable evolutionary principles, offering insights into EA behavior.