Attempt to reproduce the results from https://arxiv.org/pdf/1708.07469.pdf The merton problem is corretly solved, burger's one isnt, feel free to fill an issue The loss function and training have to be rewrite for each PDE
Just a little class to create easily samples on the domain TODO : find a way to make it more "natural"
The implementation of the architecture propose in the paper above, with modified linear layers to have a propre xavier init Using the regular pytorch layers leads prevent the model to converge properly. TODO : figure out how to initialise the bias to converge faster
The model in burger does not converge, try different hyper parameters
remove all the useless testing stuff