NOSTRA: A noise-resilient and sparse data framework for trust region based multi objective Bayesian optimization
This addresses inefficiencies in experimental resource allocation for domains like medical trials and simulations, though it appears incremental as it builds on trust region and prior knowledge integration within MOBO.
The paper tackled the problem of multi-objective Bayesian optimization (MOBO) struggling with sparse, scarce, and noisy data, common in physical experiments, by introducing NOSTRA, a framework that integrates prior uncertainty knowledge and trust regions to improve surrogate models and focus sampling, resulting in accelerated convergence and enhanced data efficiency, as demonstrated through test functions where it outperformed existing methods.
Multi-objective Bayesian optimization (MOBO) struggles with sparse (non-space-filling), scarce (limited observations) datasets affected by experimental uncertainty, where identical inputs can yield varying outputs. These challenges are common in physical and simulation experiments (e.g., randomized medical trials and, molecular dynamics simulations) and are therefore incompatible with conventional MOBO methods. As a result, experimental resources are inefficiently allocated, leading to suboptimal designs. To address this challenge, we introduce NOSTRA (Noisy and Sparse Data Trust Region-based Optimization Algorithm), a novel sampling framework that integrates prior knowledge of experimental uncertainty to construct more accurate surrogate models while employing trust regions to focus sampling on promising areas of the design space. By strategically leveraging prior information and refining search regions, NOSTRA accelerates convergence to the Pareto frontier, enhances data efficiency, and improves solution quality. Through two test functions with varying levels of experimental uncertainty, we demonstrate that NOSTRA outperforms existing methods in handling noisy, sparse, and scarce data. Specifically, we illustrate that, NOSTRA effectively prioritizes regions where samples enhance the accuracy of the identified Pareto frontier, offering a resource-efficient algorithm that is practical in scenarios with limited experimental budgets while ensuring efficient performance.