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

Authors: Michael S Yao, James Gee, Osbert Bastani

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments spanning multiple scientific domains show that Dyn AMO can be used with common optimization methods to significantly improve the diversity of proposed designs while still discovering high-quality candidates. We evaluate Dyn AMO on a set of six real-world offline MBO tasks spanning multiple scientific domains and both discrete and continuous search spaces.
Researcher Affiliation Academia 1University of Pennsylvania, Philadelphia, PA, USA. Correspondence to: Michael Yao <EMAIL>.
Pseudocode Yes Algorithm 1 (Dyn AMO). Diversity in Adversarial Modelbased Optimization
Open Source Code Yes Our custom code implementation for our experiments is made publicly available at github.com/michael-s-yao/Dyn AMO.
Open Datasets Yes All datasets used in our experiments are publicly available; offline datasets associated with Design-Bench tasks are made available by Trabucco et al. (2022). The offline dataset for the Molecule task is made available by Brown et al. (2019).
Dataset Splits No All our optimization tasks include an offline, static dataset D = {(xi, r(xi))}n i=1 of previously observed designs and their corresponding objective values. We first use D to train a task-specific forward surrogate model rθ with parameters θ according to (2).
Hardware Specification Yes All experiments were run for 10 random seeds on a single internal cluster with 8 NVIDIA RTX A6000 GPUs. Of note, all Dyn AMO experiments were run using only a single GPU.
Software Dependencies No We again use an Adam optimizer with a learning rate of η = 3 10 4 for both the VAE and the forward surrogate. sobol sequence (Sobol, 1967) using the official Py Torch quasi-random generator Sobol Engine implementation. While PyTorch is mentioned, specific version numbers for PyTorch or Adam (or its underlying library) are not provided.
Experiment Setup Yes We parameterize each forward surrogate model rθ(x) as a fully connected neural network with two hidden layers of size 2048 and Leaky ReLU activations, trained using an Adam optimizer with a learning rate of η = 0.0003 for 100 epochs. Finally, we fix the KL-divergence weighting β = 1.0, temperature hyperparameter τ = 1.0, and constraint bound W0 = 0 for all experiments to avoid overfitting Dyn AMO to any particular task or optimizer.