Meta-Learning Acquisition Functions for Transfer Learning in Bayesian\n Optimization

Michael Volpp, Lukas P. Fröhlich, Kirsten Fischer, Andreas Doerr, Stefan Falkner, Frank Hutter, Christian Daniel

arXiv (Cornell University) · 2019 · 15 citations · 0 references

DOIFull text

Open access

Concepts

Abstract

Transferring knowledge across tasks to improve data-efficiency is one of the\nopen key challenges in the field of global black-box optimization. Readily\navailable algorithms are typically designed to be universal optimizers and,\ntherefore, often suboptimal for specific tasks. We propose a novel transfer\nlearning method to obtain customized optimizers within the well-established\nframework of Bayesian optimization, allowing our algorithm to utilize the\nproven generalization capabilities of Gaussian processes. Using reinforcement\nlearning to meta-train an acquisition function (AF) on a set of related tasks,\nthe proposed method learns to extract implicit structural information and to\nexploit it for improved data-efficiency. We present experiments on a\nsimulation-to-real transfer task as well as on several synthetic functions and\non two hyperparameter search problems. The results show that our algorithm (1)\nautomatically identifies structural properties of objective functions from\navailable source tasks or simulations, (2) performs favourably in settings with\nboth scarse and abundant source data, and (3) falls back to the performance\nlevel of general AFs if no particular structure is present.\n