nlk.py file implements a non linear combination of kernels based on the following paper: https://cs.nyu.edu/~mohri/pub/nlk.pdf
simpleMKL.py file implements the simpleMKL algorithm from this paper: http://www.jmlr.org/papers/volume9/rakotomamonjy08a/rakotomamonjy08a.pdf
In practice both algorithms are not very useful. Indeed, in practise we seek to find the best combination of kernels that leads to the best performance after cross-validation, but both algorithms give the best combination of kernels possible during the training phase... because they seek the best combinations possible during the training phase...
If you ever want to achieve the best possible results on your cross-validation sets, just use a polynomial random combination of your kernels (the result is guaranteed to be a kernel). This is how my colleague https://github.com/rlespinet and I ranked 2/80 on the Kaggle for the Ecole Normale Supérieure.
PS: Go see all the kernels he implemented in C++ for the challenge here: https://github.com/rlespinet/rkernel/tree/master/lib. This guy is insane!