Project Abstract

In the field of Bayesian Optimization (BO), effectively managing computationally expensive constraints is pivotal, despite advancements in traditional acquisition functions (AFs) such as Expected Improvement (EI), Probability of Improvement (PI), and Predictive Entropy Search (PES). These functions are essential for guiding the selection of parameter values for future evaluations. Although their approach to balancing exploration and exploitation varies, they often struggle to incorporate constraints effectively in various scenarios. This project aims to reimplement and potentially enhance the performance of selected state-of-the-art AFs to better handle different constraints. The initial phase involves a detailed analysis and review of existing constraint-specific AFs to establish a robust baseline. Subsequently, we propose to develop and evaluate novel AFs with the objective of surpassing this baseline in terms of both efficiency and performance. By employing advanced data-driven techniques and optimizing AF integration, the project seeks to demonstrate potential improvements in optimizing complex cost functions under stringent computational constraints. The anticipated outcome is an optimized set of strategies that enhance decision-making in scenarios where computational resources are limited, thereby contributing significantly to fields that demand efficient optimization solutions. This approach aims to provide a comprehensive understanding of how AFs can be modified to incorporate various constraint types, enhancing BO’s utility across a broader range of applications.

Keywords: Bayesian Optimization and Acquisition Functions, Optimization in Constrained Environments, Computational Efficiency

 Conference Details

Session: Presentation Stream 5 at Presentation Slot 4

Location: GH011 at Tuesday 7th 13:30 – 17:00

Markers: Nader Al Khatib (GTA), Tom Owen

Course: MSc Data Science, Masters PG

Future Plans: I’m looking for work