Azhir Mahmood  🚀 🤖 🧬 🧠 🪐

A recent thread about practical use of GPs reminded me that I haven’t posted about the 5 BayesOpt-related papers the Adaptive Experimentation team at Facebook is presenting at NeurIPS 2020 this year. The methods from these papers are all being applied to real-world problems.

How effective is the novel one-shot formulation, and have you applied it to real world datasets?

Besides examples in the paper, one-shot KG performs better than qNEI, which tends to perform better than implementations f EI we've tested. KG is particularly useful for multi-fidelity optimization, for which we see great performance on AutoML and simulation optimization tasks.