Explainable Meta Bayesian Optimization with Human Feedback for Scientific Applications like Fusion Energy (Papers Track)

Ricardo Luna Gutierrez (Hewlett Packard Enterprise); Sahand Ghorbanpour (Hewlett Packard Enterprise); Vineet Gundecha (Hewlett Packard Enterpise); Rahman Ejaz (University of Rochester); Varchas Gopalaswamy (University of Rochester); Riccardo Betti (University of Rochester); Avisek Naug (Hewlett Packard Enterprise); Desik Rengarajan (Hewlett Packard Labs); Ashwin Ramesh Babu (Hewlett Packard Enterprise Labs); Paolo Faraboschi (Hewlett Packard Enterprise); Soumyendu Sarkar (Hewlett Packard Enterprise)

Paper PDF Slides PDF Poster File NeurIPS 2024 Recorded Talk Cite
Power & Energy Reinforcement Learning

Abstract

We introduce Meta Bayesian Optimization with Human Feedback (MBO-HF), which integrates Meta-Learning and expert preferences to enhance BO. MBO-HF employs Transformer Neural Processes (TNPs) to create a meta-learned surrogate model and a human-informed acquisition function (AF) to suggest and explain proposed candidate experiments. MBO-HF outperforms current methods in optimizing various scientific experiments and benchmarks in simulation, including the energy yield of the inertial confinement fusion (ICF), practical molecular optimization (PMO), and critical temperature maximization for superconducting materials.