Austin Tripp's website

Behaviour-based evaluations in Bayesian optimization

Austin Tripp — Sat, 21 Feb 2026 22:00:00 GMT

I think the Bayesian optimization (BO) research community needs to change its evaluation practices. In a previous post I explained how I think users should generally bring their own model to BO and try to calibrate it to their beliefs/expectations. In this post, I explain why existing evaluation practices are not well-suited to this stance on BO, and outline a different approach to BO evaluation based on algorithm behaviour.

Rebranding BO away from "black-box" and towards "model-based"

Austin Tripp — Mon, 16 Feb 2026 15:00:00 GMT

In a recent blog post (link) I described my "model-centric" view of Bayesian optimization (BO), essentially arguing that the model is the most important component of BO and BO users (and researchers) should do more to get it right. Under the assumption that the reader broadly agrees with the content of that post (or at least thinks this view of BO is one of several valid views), here I want to argue that the BO community should re-brand itself more towards model-based optimization and away from black-box optimization.

Clarifying noise vs model misspecification in Gaussian Process models (and its importance in BO)

Austin Tripp — Sat, 14 Feb 2026 23:00:00 GMT

The most common kind of Gaussian process (GP) model is:

$$f \sim GP\left(\mu(\cdot), k(\cdot, \cdot)\right) $$

$$ y \sim N(f(x), \sigma^2) $$

Tyler Cowen's book on talent

Austin Tripp — Fri, 13 Feb 2026 00:00:00 GMT

Cross-posting my summary of the book "Talent: How to Identify Energizers, Creatives, and Winners Around the World" from my GitHub: link.

We are underselling the modularity of Bayesian optimization

Austin Tripp — Mon, 09 Feb 2026 00:00:00 GMT

I think the Bayesian optimization (BO) community is vastly underselling one of BO's most practically appealing features: modularity. A "modular" algorithm is one composed of individual parts which are largely exchangeable. I think BO is extremely modular, and this is something users want in practice, but BO papers rarely describe it in this way. This post outlines these points in more detail and ends with some recommendations.

Why I don't care about toy benchmarks in BO

Austin Tripp — Sun, 08 Feb 2026 22:00:00 GMT

A lot of Bayesian optimization (BO) papers include experiments on "toy" functions. Examples of such toy functions are Ackley, Rosenbrock, Hartmann, and Branin. I basically ignore these experiments when I read papers and skip to the next section. When I review papers (and therefore cannot simply skip sections), I find my eyes glazing over and feel bored: I don't care about these functions, don't care about the results, and don't really understand why the authors ran this experiment.

For a while I assumed every BO researcher feels this way, but a bunch of recent conversations have made it clear to me that many people do care. So, in this post I explain why I don't care at all about benchmarks on toy functions (and why you shouldn't either).

We have forgotten about utility functions in BO (whoops!)

Austin Tripp — Tue, 27 Jan 2026 00:00:00 GMT

Bayesian decision theory is one of the best justifications for BO- particularly for myopic acquisition functions like expected improvement. However, these acquisition functions are only "optimal" if one's utility function is $u(y) = y$ (the identity function). Have BO researchers (and BO users) basically forgotten to swap this out for "real" utility functions in practice? In this post I argue that we have basically overlooked this detail (to our own detriment). It's not too hard to fix, but unfortunately the $u(y) = y$ assumption is actually quite deeply embedded, and completely removing it will make things more complicated. Ultimately, despite the difficulty, I think we should do it anyway.

My model-centric view of Bayesian optimization

Austin Tripp — Sun, 25 Jan 2026 22:00:00 GMT

A lot of conversations at NeurIPS this year made me think that I view the role of the surrogate model in Bayesian optimization a bit differently than many other researchers in the field, and this profoundly impacts my view of many other aspects of BO. Therefore, the purpose of this post is to explain my view and contrast it with what I believe is the more mainstream view.

Predictions for ML/AI in 2026 (and 2025 predictions re-visited).

Austin Tripp — Mon, 05 Jan 2026 01:00:00 GMT

Last year I made a bunch of predictions about ML, and since 2025 is over it's time to grade myself and repeat this exercise.

New Year's Resolutions for 2026 (and scoring 2025 resolutions)

Austin Tripp — Mon, 05 Jan 2026 00:00:00 GMT

Happy 2026! Here is my self-assessment on last year's goals and my goals for this year (2026).