A paper co-authored by a researcher at MI-6 Ltd., entitled “Improved Regret Analysis for Parallel Gaussian Process Bandit Optimization”, has been accepted as a Spotlight paper at NeurIPS 2026, one of the world’s leading international conferences in machine learning.
The NeurIPS 2026 Main Track received 30,709 valid submissions, of which 7,900 papers (25.7%) were accepted. Among the accepted papers, 7,496 were selected as Posters, 292 as Spotlights, and 112 as Orals. Spotlight papers therefore accounted for approximately 1% of all submissions.
Research Overview
This study presents a new theoretical analysis of performance guarantees for parallel Bayesian optimization, in which multiple candidate points are evaluated concurrently.
Bayesian optimization is a method for efficiently optimising an unknown objective function. It aims to identify promising conditions while reducing the number of costly experiments or simulations required. In settings where multiple experiments can be conducted simultaneously, parallel optimisation — in which several candidate points are selected and evaluated as a batch — is particularly important.
Previous theoretical analyses of representative parallel Bayesian optimization algorithms have suffered from regret bounds that deteriorate multiplicatively with the degree of parallelism, or batch size. To avoid this deterioration, existing analyses have required an initial phase of uncertainty-based sampling before optimisation begins.
In this work, we analyse GP Batched Thompson Sampling (GP-BTS) and show that a regret bound without such multiplicative dependence on the batch size can be obtained without requiring this initial sampling phase. We further establish an improved regret bound for the noiseless observation setting compared with the corresponding bound in the noisy setting.
These results advance the theoretical understanding of Bayesian optimization in settings where multiple experiments or evaluations are carried out in parallel.
Materials Development and Bayesian Optimisation
Materials development often involves experiments and simulations that are both time-consuming and costly. At the same time, there are many situations in which multiple candidate conditions can be evaluated in parallel.
Materials Development and Bayesian Optimisation
Materials development often involves experiments and simulations that are both time-consuming and costly. At the same time, there are many situations in which multiple candidate conditions can be evaluated in parallel.
At MI-6, we regard Bayesian optimisation, which uses machine learning to recommend the experimental conditions that should be investigated next, as one of the key enabling technologies for accelerating and improving the efficiency of materials development.
Through theoretical research into parallel optimisation such as this study, we aim to deepen our understanding of the performance and characteristics of optimisation algorithms, while building the technological foundations needed to enable more efficient and reliable exploration in real-world materials research and development.
Paper Information
Title
Improved Regret Analysis for Parallel Gaussian Process Bandit Optimization
Authors
Shion Takeno (Nagoya University)
Shogo Iwazaki (MI-6 Ltd.)
Conference
NeurIPS 2026 Main Track (Spotlight)
Paper
arXiv: 2608.16492
https://arxiv.org/abs/2608.16492
MI-6 will continue to pursue research and development at the intersection of materials science and machine learning, contributing to the advancement of machine-learning technologies that support and accelerate materials development.
