Gaussian Processes offer a versatile tool for modelling the underlying behaviour of partially observed systems. However it can be challenging to find an appropriate choice of kernel for a given problem. One approach is to construct a kernel by combining a broad variety of simpler kernels, via a sequence of layers which resembles a neural network. The technique was originally presented in a paper entitled 'Differentiable Compositional Kernel Learning for Gaussian Processes' by Sun et al. The implementation in this repository is designed to be used as part of the GPflow package.
The package can be installed by cloning the repository and running the following commands from the root folder:
pip install -r requirements.txt
python setup.py install
Once installed, you should be able to run the script '/demo/run_hybrid_kernels.py', which produces the figure below.