Core Concepts
Proposing new geometric metrics to enhance the performance evaluation of Bayesian optimization algorithms.
Abstract
The content discusses the introduction of geometric metrics in Bayesian optimization to address limitations in conventional regret-based metrics. It introduces precision, recall, average degree, and average distance metrics to evaluate the geometry of query points and global solutions. The parameter-free forms of these metrics are also proposed to eliminate the need for additional parameters. Numerical analyses on metric values over iterations and Spearman's rank correlation coefficients between metrics are presented.
Structure:
Introduction to Bayesian Optimization
Limitations of Regret-Based Metrics
Proposal of Geometric Metrics
Parameter-Free Forms of Metrics
Data Extraction and Analysis on Benchmark Functions
Discussion on Related Work and Limitations
Stats
Bayesian optimization is a principled strategy with a probabilistic regression model.
Proposed geometric metrics include precision, recall, average degree, and average distance.
Parameter-free forms of these metrics are suggested for ease of use.