This repository contains the code used to produce the results in the paper Efficient Rollout Strategies for Bayesian Optimization. This code is largely based off the Metric Optimization Engine (MOE) open-source Bayesian optimization library, albeit simplified in many areas. We note that this code is not meant to be a fully-fledged Bayesian optimization software package (e.g., it does not support categorical variables), and is primarily intended to illustrate the more important concepts of our paper.
To keep things simple, we recommend the latest version of Anaconda. Otherwise, our requirements are:
- Python >= 3.6
- The latest version of numpy, scipy, jupyter and torch.
Run python setup.py install in the command line.
The lookahead/ directory is subdivided as follows:
lookahead/acquisitions/contains implementations of all acquisition functions. This includes expected improvement, upper confidence bound, knowledge gradient, rollout of EI, and policy search.lookahead/model/contains our GP implementation. By default, this implementation uses a constant mean function and the Matern 5/2 kernel.lookahead/runners/contains the BO runners, which are invoked to run a full BO loop.lookahead/test_problems/contains implementations of synthetic functions, which are used to benchmark BO acquisition functions.
We have a few simple demos in the demos/ folder, which you should run to get started.