Gaussian-Process Surrogate Optimisation

This is a new implementation of IMGPO which brings several improvements:

• Correct update of GP hyperparameters, and subsequent update of UCBs;
• GP trained in normalised space, to honour isotropic covariance assumptions;
• Correct count of GP- and non-GP-based samples;
• Decoupling of GP surrogate, partition tree and optimisation logic;
• Object orientation and events for better clarity and interfacing;
• Serialisation, configuration and numerous methods for analysis.

Installation

You will need Deck to use this toolbox. If it is not already installed, navigate to the folder where you want to put it (eg /Users/me/Documents/MATLAB), and type:

git clone https://github.com/sheljohn/deck.git

This will create a folder deck/; add it to your path, and type dk_startup from the Matlab console. If you need to use this toolbox frequently, add these last commands to your startup.m.

You will also need to be able to compile Mex files; make sure Matlab is set up properly (if you are on OSX, this involves installing the Command-Line Tools and Xcode, you might also want to use Homebrew to install gcc and/or up-to-date versions of clang). For a quick verification you can type mex -setup from the console.

Once you have Deck installed, and that Mex is setup up properly, install GPSO if your folder of chocie with:

git clone https://github.com/sheljohn/gpso.git

This will create a folder gpso/; add it to your path, and type gpml_compile from the Matlab console.

Usage

Running optimisation

Create, configure, run, resume

Extras

Using events, serialising, exporting the tree

Examples

Can be accessed through gpso_example.*

License

GNU AGPL v3