Standard Acquisition Is Sufficient for Asynchronous Bayesian Optimization

A team of researchers including Ben Riegler, James Odgers, and Vincent Fortuin has published a groundbreaking paper titled 'Standard Acquisition Is Sufficient for Asynchronous Bayesian Optimization' that challenges fundamental assumptions in the field of parallel optimization. The work demonstrates that standard acquisition functions like Upper Confidence Bound (UCB) can achieve theoretical guarantees essentially equivalent to those of sequential Thompson sampling in asynchronous settings, where multiple experiments run in parallel with varying completion times. This finding contradicts the prevailing wisdom that specialized, complex methods are necessary to prevent redundant queries in such environments.

The researchers conducted a conceptual analysis revealing that existing asynchronous Bayesian optimization methods have been neglecting intermediate posterior updates—the continuous refinement of probability distributions as new data arrives from parallel experiments. They found that these updates are generally sufficient to avoid the redundant queries that complex methods try to prevent through artificial diversity enforcement. In fact, their investigation shows that diversity-enforcing methods can actually over-explore in asynchronous settings, reducing their performance compared to simpler approaches.

Extensive experiments across synthetic and real-world tasks consistently demonstrated that simple standard acquisition functions match or outperform purpose-built asynchronous methods. The paper provides both theoretical guarantees and empirical evidence that challenges the need for complex algorithmic solutions to what turns out to be a simpler problem than previously thought. This work has significant implications for how researchers and practitioners approach parallel optimization problems in fields ranging from drug discovery to hyperparameter tuning.