Generative Adversarial Model-Based Optimization via Source Critic Regularization
This work addresses a critical bottleneck in offline optimization for domains like protein design and robotics, where evaluating the true objective is expensive, though it appears incremental as it builds on standard Bayesian optimization.
The paper tackles the problem of inaccurate surrogate model predictions in offline model-based optimization by proposing a generative adversarial framework with adaptive source critic regularization (aSCR), which constrains optimization to reliable regions and outperforms existing methods on generative design tasks.
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