Generative Adversarial Model-Based Optimization via Source Critic Regularization.

Area of Science:

• Computational Biology
• Machine Learning
• Optimization

Background:

• Offline model-based optimization uses surrogate models for expensive objective functions.
• Inaccurate surrogate predictions hinder optimization performance in fields like protein design and robotics.
• Existing methods struggle with reliability in offline optimization trajectories.

Purpose of the Study:

• To introduce a novel framework, generative adversarial model-based optimization with adaptive source critic regularization (aSCR), for reliable offline optimization.
• To constrain optimization to regions where the surrogate model is accurate.
• To enhance performance in computationally expensive design tasks.

Main Methods:

• Proposed generative adversarial model-based optimization (GABO) framework.
• Introduced adaptive source critic regularization (aSCR) as a task- and optimizer-agnostic constraint.
• Developed a tractable algorithm for dynamic constraint adjustment.
• Integrated aSCR with standard Bayesian optimization.

Main Results:

• aSCR effectively constrains optimization to reliable surrogate model regions.
• The proposed algorithm dynamically adjusts regularization strength.
• Leveraging aSCR with Bayesian optimization outperformed existing methods on offline generative design tasks.
• Demonstrated improved performance across a suite of tasks.

Conclusions:

• Generative adversarial model-based optimization with aSCR provides a robust solution for offline optimization challenges.
• The framework enhances reliability by ensuring optimization within trusted surrogate model predictions.
• aSCR offers a significant advancement for computationally expensive generative design problems.