Gaussian processes for decision-making
Room: Kilburn Building, Lecture Theatre 1.1
9:30-10:00 Arrivals (includes coffee)
10:00-10:10 Welcome
10:10-11:10 Experimental Design in the Age of GenAI
11:10-12:10 TBC
12:10-13:30 Lunch
13:30-14:30 Global Acquisition Optimization for Structured Graph Bayesian Optimization
14:30-15:00 Tea Break
15:00-16:00 TBC
Abstracts
Experimental Design in the Age of GenAI
Abstract
Generative AI stands to revolutionize how experiments are conceived, orchestrated, and iterated. This talk outlines a general framework that can harness generative AI within experimental design loops, then examines the widely adopted approach of Latent Space Bayesian Optimization. Our proposed methodology diverges from conventional frameworks by decoupling the discriminative (surrogate) and generative models, rather than tightly integrating Gaussian Processes with Variational Autoencoders. Through independent training and a straightforward Bayesian update, we enable an efficient sampling strategy that identifies candidate structures with high scores.
TBC
Abstract
Global Acquisition Optimization for Structured Graph Bayesian Optimization
Abstract
Bayesian optimization over graph structured domains has emerged as a powerful paradigm for tasks such as molecular design and neural architecture search (NAS). While noticeable progress has been made on constructing expressive graph based Gaussian process (GP) surrogates, acquisition function optimization central to BO performance remains underexplored in discrete, combinatorial graph spaces, particularly under structural constraints. This talk addresses the challenge of global acquisition optimization in such settings. We present recent work that encodes both the graph search space and shortest path graph kernels via mixed integer programming (MIP), enabling acquisition functions such as LCB to be globally optimized. Empirical results on molecular design and NAS benchmarks demonstrate the effectiveness of principled acquisition optimization in graph structured BO.