ensemble model training

[aapohaavisto]

Hi,
To train an ensemble model, one woud need to be ble to get different representitive points each time. is there any way one can pick the sparse representitive points self?

[mcaroba]

I'm doing some tests on ensemble predictions within GAP at the moment. If these are promising, it will likely be implemented as a training feature in the near future. The computationally efficient way of doing ensemble predictions actually fixes the sparse points, so that the kernels only have to be computed once for all models and it's the alphas (the fitting coefficients) that change between models. One can fix the sparse set by using the sparse_method=file mode in gap_fit (which needs a sparse_file=your_descriptors.dat). Currently, the only way of doing this is by manually training the models and manually adding up the predictions (and computing the statistical uncertainty). This is what my tests are looking at. I can tell you more in person.