Diversity By Design: Leveraging Distribution Matching for Offline Model-Based Optimization

Michael S Yao, James Gee, Osbert Bastani

TL;DR: We introduce a method to propose both high-quality and diverse designs in a wide variety of offline model-based optimization tasks.

Abstract: The goal of offline model-based optimization (MBO) is to propose new designs that maximize a reward function given only an offline dataset. However, an important desiderata is to also propose a *diverse* set of final candidates that capture many optimal and near-optimal design configurations. We propose **D**iversit**y** I**n** **A**dversarial **M**odel-based **O**ptimization (**DynAMO**) as a novel method to introduce design diversity as an explicit objective into any MBO problem. Our key insight is to formulate diversity as a *distribution matching problem* where the distribution of generated designs captures the inherent diversity contained within the offline dataset. Extensive experiments spanning multiple scientific domains show that DynAMO can be used with common optimization methods to significantly improve the diversity of proposed designs while still discovering high-quality candidates.

Lay Summary:

When scientists try to design new things (e.g., medicines, materials, or robots), they often rely on computer models trained on previously collected data to suggest promising candidate designs. This is useful because testing each design in real life can be very expensive or slow. However, these computer models are not always accurate. Furthermore, existing methods often 'play it safe' and suggest very similar designs, potentially missing out on other great options.

Our research introduces **DynAMO**, a new tool that helps computers suggest a greater *diversity* of potentially high-quality designs. Instead of just looking for the single 'best' solution, DynAMO is an algorithm that explores **many different kinds of good ideas** by mimicking the diversity seen in past successful designs. It also enforces constraints to ensure the proposed designs are realistic and make sense.

We tested DynAMO on several real-world design problems, such as engineering DNA sequences and designing new robots, and found that DynAMO suggests more diverse designs without sacrificing their quality.

Why does this matter? In real-world innovation, having a broader set of strong options increases the chances of discovering something truly groundbreaking. DynAMO makes that exploration faster and more trustworthy.

Link To Code: https://github.com/michael-s-yao/DynAMO

Primary Area: Optimization->Zero-order and Black-box Optimization

Keywords: Offline Optimization, AI4Science, Generative Design, Model-Based Optimization

Submission Number: 11461