sharpoo7 / gaussian-process-regression

Simple library with a basic no-frills implementation of GPR using Eigen. Basic support for multidimensional outputs.

Gaussian Process Regression

This is a library implementing standard GP-regression. Currently it uses only the squared exponential covariance function but can easily be modified for differnt covariance functions.

This is a template library - no linking required. Just make sure your code can find GaussianProcessRegression.h and you should be good to go.

Dependencies

The version here is packaged as a catkin package for easy use with ROS.

It depends on Eigen3.

Basic Usage

// input_dim and output_dim are integers specifying the dimensionality of input and output data respectively
GaussianProcessRegression<float> myGPR(input_dim, output_dim);
// the hyperparameters are float values specifying length-scale, signal variance and observation noise variance
myGPR.SetHyperParams(length_scale, sigma_f, sigma_n);
Eigen::VectorXf train_input(input_dim);
Eigen::VectorXf train_output(output_dim);
// add some training data
int n_train = 100;
for(int k=0; k<n_train; k++){
	train_input.setRandom();
	train_output.setRandom();
	myGPR.AddTrainingData(train_input,train_output)
}
// get output at a testing point
Eigen::VectorXf test_input(input_dim);
Eigen::VectorXf test_output(output_dim);
test_input.setRandom();
test_output = myGPR.DoRegression(test_input);

Notes on multidimensional outputs

As seen in the example above, the library supports multiple outputs, but in a kind of "dumb" way. The same covariance function and inputs are used to model all the outputs.

TODO

• Add possibliity to use user-supplied covariance function.
• Add possibility to get variance information