FDDEDO: A Novel Federated Data-Driven Evolutionary Dynamic Optimization Framework

What the paper says

In the real world, many optimization problems involve time-varying and computationally expensive objective functions. Data-driven surrogate-assisted evolutionary algorithms (SAEAs) are considered promising approaches for solving these problems. However, data-driven SAEAs may face problems in terms of privacy protection and data security when real data are stored in a distributed form across different client devices. Furthermore, the heterogeneity of data stored across different clients further complicates optimization efforts. To address the aforementioned problems, this article proposes a novel federated data-driven evolutionary dynamic optimization framework called FDDEDO. Specifically, to enhance server-side aggregation capabilities, we propose a cosine distance-based surrogate aggregation method that improves the performance of global radial basis function network (RBFN) surrogate through robust RBFN center matching. To cope with environmental changes under limited evaluation budget, a gradient-based client-side surrogate meta-training algorithm is proposed to generate efficient initial local surrogates embedded with prior knowledge for new environments by dynamically learning transfer patterns among historical environments. Meanwhile, to strengthen privacy protection while adapting to heterogeneous data, a client-side surrogate adaptation process based on differential privacy (DP) mechanism is designed. By introducing (ϵ, δ)-DP and combining it with the proximal term used to balance local personalization with global consistency, it achieves effective privacy-preserving fine-tuning for local surrogate under limited samples. Experimental results on benchmark problems under homogeneous and heterogeneous federated settings demonstrate that FDDEDO exhibits significant superiority in overall optimization performance, with all its core components operating effectively.