Generative Adversarial Model-Based Optimization via Source Critic Regularization

TL;DR

This paper introduces a novel offline optimization framework that uses adversarial training with adaptive regularization to improve the reliability of surrogate models in expensive evaluation scenarios like protein design and robotics.

Contribution

It proposes a source critic regularization method that dynamically constrains optimization trajectories to reliable surrogate regions, enhancing performance over existing methods.

Findings

Outperforms existing offline optimization techniques

Effectively constrains optimization to reliable surrogate regions

Applicable across diverse 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