Machine Learning for Hidden Physics and Partial Differential Equations
- Using pyGPs (has limitations)
- with the Backward Euler scheme
- Non-homogeneous problem
- Using 2D kernel
The Python package mlhiphy can be installed in the traditional way
- Standard mode::
python3 -m pip install .
- Development mode::
python3 -m pip install --user -e .
In https://github.com/Slowpuncher24/mlhiphy_v2 you can find a considerably improved version of mlhiphy. It features:
- A much more efficient and stable implementation of the negative log-likelihood. This vastly improves the algorithm, as the optimization of the negative log-likelihood is at its center. This was done by utilizing the block matrix structure of the covariance matrix and by using the Cholesky decomposition.
- The inference of up to four hidden parameters in three dimensions, as opposed to mainly one hidden parameter in two dimensions in mlhiphy (respectively counting the temporal dimension as one).
- An alternative implementation of the negative log-likelihood for the noise-free case, where we can optimize over one hyperparameter less (the signal variance can be written in terms of other values).
- The implementation and tests of using the Matérn-5/2-kernel, which is the most promising alternative to the SE kernel.
- The implementation and tests of a viable alternative to the Nelder-Mead optimization algorithm, namely a variant of the nonlinear conjugate gradient method (it is scipy's implementation up to a minor tweak to the line search algorithm).