kirthevasank / mf-gp-ucb

Multi-fidelity Gaussian Process Bandit Optimisation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

mf-gp-ucb

This is a Matlab implementation of the Multi-fidelity Gaussian Process Upper Confidence Bound algorithm for Bayesian optimisation with multi-fidelity approximations. For more details please read our paper (below).

Download

You can download the code from github.

$ git clone https://github.com/kirthevasank/mf-gp-ucb 

Installation & Getting Started

  • Simply execute the following where path-to-mf-gp-ucb is the download directory.
>> addpath(genpath('path-to-mf-gp-ucb'))
  • The demos directory has two examples on how to use this library.
  • demo_short.m: Easy set up and uses all default configurations.
  • demo_long.m: This demonstrates how to set all parameters individually in case you want to customise them. (TBD)
  • To be able to use the software, you will need to construct an mfFunction Object. This is quite straightforward. See the examples in demos/functions.

Some notes

  • We choose the GP hyper-parameters every 25 iterations via marginal likelihood maximisation for each GP. The chosen values are printed out.
  • We report progress on the optimisation every 10 iterations. We report the cost incurred, the number of queries at each fidelity and the maximum value found so far.

Citation

If you use this library in your academic work please cite our NIPS 2016 paper: "Gaussian Process Bandit Optimisation with Multi-fidelity Evaluations", Kirthevasan Kandasamy, Gautam Dasarathy, Junier Oliva, Jeff Schneider, Barnabas Poczos.

We use DiRect to optimise the acquisition function. The implementation was taken from Dan Finkel (2004).

License

This software is released under the MIT license. For more details, please refer LICENSE.txt.

"Copyright 2015 Kirthevasan Kandasamy"

About

Multi-fidelity Gaussian Process Bandit Optimisation

License:MIT License


Languages

Language:MATLAB 100.0%