Transfer learning outperforms Gaussian processes in multi-fidelity Bayesian optimization

A new research paper explores the use of transfer learning architectures as a core component for multi-fidelity Bayesian optimization (MFBO). The study benchmarks eleven transfer-learning surrogates against traditional Gaussian processes (GPs) across various tasks, including synthetic functions and real-world chemistry and materials problems. Results indicate that transfer learning significantly outperforms GPs on molecular and materials problems, achieving better solutions with less computation, making it the preferred surrogate for MFBO in these domains. AI

IMPACT This research could lead to more efficient AI-driven discovery in chemistry and materials science by improving optimization techniques.

RANK_REASON Research paper introducing and evaluating a new methodology for Bayesian optimization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Bayesian optimization
• Gaussian processes
• transfer learning

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →