Robust Guided Diffusion for Offline Black-Box Optimization

TL;DR

Offline black-box optimization (BBO) seeks high-performing designs from offline data. The paper introduces Robust Guided Diffusion for Offline Black-box Optimization (RGD), which combines explicit proxy guidance with robust proxy-free diffusion through two modules: proxy-enhanced sampling and diffusion-based proxy refinement. Proxy-enhanced sampling optimizes a strength parameter $oldsymbol{ extomega}$ to balance conditioning and sample diversity, while diffusion-based proxy refinement regularizes the proxy by aligning it with a diffusion-derived distribution on adversarial samples via KL-divergence and bi-level optimization. Empirically, RGD achieves state-of-the-art results on Design-Bench tasks, with ablations confirming the value of both components and analyses showing robustness to hyperparameters; code is available at the provided GitHub link.

Abstract

Offline black-box optimization aims to maximize a black-box function using an offline dataset of designs and their measured properties. Two main approaches have emerged: the forward approach, which learns a mapping from input to its value, thereby acting as a proxy to guide optimization, and the inverse approach, which learns a mapping from value to input for conditional generation. (a) Although proxy-free~(classifier-free) diffusion shows promise in robustly modeling the inverse mapping, it lacks explicit guidance from proxies, essential for generating high-performance samples beyond the training distribution. Therefore, we propose \textit{proxy-enhanced sampling} which utilizes the explicit guidance from a trained proxy to bolster proxy-free diffusion with enhanced sampling control. (b) Yet, the trained proxy is susceptible to out-of-distribution issues. To address this, we devise the module \textit{diffusion-based proxy refinement}, which seamlessly integrates insights from proxy-free diffusion back into the proxy for refinement. To sum up, we propose \textit{\textbf{R}obust \textbf{G}uided \textbf{D}iffusion for Offline Black-box Optimization}~(\textbf{RGD}), combining the advantages of proxy~(explicit guidance) and proxy-free diffusion~(robustness) for effective conditional generation. RGD achieves state-of-the-art results on various design-bench tasks, underscoring its efficacy. Our code is at https://github.com/GGchen1997/RGD.

Figures (8)

• Figure 1: Motivation of explicit proxy guidance.
• Figure 2: Overview of RGD: Module (a) incorporates proxy guidance into proxy-free diffusion to enable enhanced sampling control; Module (b) integrates insights from proxy-free diffusion back into the proxy for refinement.
• Figure 3: This adjustment of $\omega$ effectively balances between generating novel solutions and honing in on high-quality ones during sampling.
• Figure 4: The proxy distribution overestimates the ground truth, while the diffusion distribution closely aligns with it, demonstrating its robustness.
• Figure 5: The ratio of the performance of our RGD method with $T$to the performance with $T=1000$.