Learning Evolution via Optimization Knowledge Adaptation

A research team led by Chao Wang has published a paper introducing the Optimization Knowledge Adaptation Evolutionary Model (OKAEM), a novel framework designed to unify two critical but often separate goals in evolutionary algorithms (EAs): knowledge transfer and online adaptation. Current methods for sequential transfer optimization often fail to fully utilize prior knowledge, while adaptive strategies are limited to tweaking specific operators. OKAEM addresses this by using attention mechanisms to parameterize evolutionary operators, creating a single, learnable model capable of both integrating extensive historical knowledge and dynamically updating itself in real-time based on new optimization data.

Experimental results confirm OKAEM's significant advantages. The model outperformed state-of-the-art sequential transfer methods across 12 different transfer scenarios through its pre-training phase. Furthermore, in settings with no prior knowledge, its self-tuning mechanism surpassed advanced learnable EAs. Beyond benchmark tests, the team demonstrated OKAEM's practical utility by applying it to prompt tuning for vision-language models. Ablation studies validated the necessity of its learnable components, and visualization analyses revealed the model's unique capacity to autonomously discover underlying, interpretable evolutionary principles, moving beyond simple black-box optimization. The work has been accepted for publication in the IEEE Transactions on Pattern Analysis and Machine Intelligence.