Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Adversarial Model-Based Optimization via Source Critic Regularization

Michael S. Yao, Yimeng Zeng, Hamsa Bastani, Jacob Gardner, James C. Gee, Osbert Bastani

TL;DR

This work proposes adaptive source critic regularization (aSCR)-a task- and optimizer- agnostic framework for constraining the optimization trajectory to regions of the design space where the surrogate function is reliable, and shows how leveraging aSCR with standard Bayesian optimization outperforms existing methods on a suite of offline generative design tasks.

Abstract

Offline model-based optimization seeks to optimize against a learned surrogate model without querying the true oracle objective function during optimization. Such tasks are commonly encountered in protein design, robotics, and clinical medicine where evaluating the oracle function is prohibitively expensive. However, inaccurate surrogate model predictions are frequently encountered along offline optimization trajectories. To address this limitation, we propose generative adversarial model-based optimization using adaptive source critic regularization (aSCR) -- a task- and optimizer- agnostic framework for constraining the optimization trajectory to regions of the design space where the surrogate function is reliable. We propose a computationally tractable algorithm to dynamically adjust the strength of this constraint, and show how leveraging aSCR with standard Bayesian optimization outperforms existing methods on a suite of offline generative design tasks. Our code is available at https://github.com/michael-s-yao/gabo

Generative Adversarial Model-Based Optimization via Source Critic Regularization

TL;DR

This work proposes adaptive source critic regularization (aSCR)-a task- and optimizer- agnostic framework for constraining the optimization trajectory to regions of the design space where the surrogate function is reliable, and shows how leveraging aSCR with standard Bayesian optimization outperforms existing methods on a suite of offline generative design tasks.

Abstract

Offline model-based optimization seeks to optimize against a learned surrogate model without querying the true oracle objective function during optimization. Such tasks are commonly encountered in protein design, robotics, and clinical medicine where evaluating the oracle function is prohibitively expensive. However, inaccurate surrogate model predictions are frequently encountered along offline optimization trajectories. To address this limitation, we propose generative adversarial model-based optimization using adaptive source critic regularization (aSCR) -- a task- and optimizer- agnostic framework for constraining the optimization trajectory to regions of the design space where the surrogate function is reliable. We propose a computationally tractable algorithm to dynamically adjust the strength of this constraint, and show how leveraging aSCR with standard Bayesian optimization outperforms existing methods on a suite of offline generative design tasks. Our code is available at https://github.com/michael-s-yao/gabo
Paper Structure (27 sections, 13 equations, 6 figures, 11 tables, 3 algorithms)

This paper contains 27 sections, 13 equations, 6 figures, 11 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background
  4. Offline Model-Based Optimization
  5. Optimization Over Latent Spaces
  6. Wasserstein Metric Between Probability Distributions
  7. A Framework for Generative Adversarial Optimization
  8. Constrained Optimization Formulation
  9. Dual Formulation
  10. Computing the Lagrange Multiplier Lambda
  11. Overall Algorithm
  12. Experimental Evaluation
  13. Datasets and Tasks
  14. Policy Optimization and Evaluation
  15. Main Results
  16. ...and 12 more sections

Figures (6)

  • Figure 1: Naïve offline model-based optimization (MBO) coms, which optimizes against a learned surrogate model $f_{\theta}$ trained on a fixed dataset $\mathcal{D}_n=\{(\mathbf{x}_i, y_i)\}_{i=1}^n$ (shaded region) without access to the true oracle $f$, often yields candidate designs $\mathbf{x}^*$ (i.e., diamond) that score poorly using the true oracle (i.e., cross). Our method (aSCR) constrains optimization trajectories to avoid these extrapolated points, instead proposing 'in-distribution' designs (i.e., star).
  • Figure 2: Penalized LogP Score Maximization Sample Candidate Designs$\quad$ (Left) The molecule with the highest penalized LogP score of 11.3 in the offline dataset. Separately, we show the 100th percentile candidate molecules according to the surrogate objective generated from (Middle) vanilla BO-qEI and (Right) GABO. Teal- (white-) colored atoms are carbon (hydrogen). Non-hydrocarbon atoms are underlined in the SMILES smiles string representations of the molecules.
  • Figure : Adaptive Source Critic Regularization (SCR)
  • Figure B1: Distribution of Oracle Penalized LogP Scores We plot the distribution of oracle scores for the top 128 surrogate model-ranked designs in black, and the distribution for all 2,048 generated designs in light gray for each of the offline model-based optimization methods assessed in our work across 10 random seeds. While GABO and BO-qEI have similar distributions, GABO is able to more reliably rank top-performing designs higher, such that these designs can be identified even under limited oracle query budgets.
  • Figure B2: 100th Percentile Oracle Scores versus $k$-Shot Oracle Budget Size$\quad$ We plot the 100th percentile oracle Penalized LogP score averaged across 10 random seeds as a function of the number of allowed oracle calls $k$.
  • ...and 1 more figures