TODO for neural network approximation of ODE

[mikeaalv]

The network method has worked for comparably small ODE set (dimension=4), and there is obvious trouble in making it work for larger ODE set. Here I list a few things that has been tried or might work (but haven't been tried).

• rennet

• optimizer scheduler

• automatic hyper-parameter searching (such as optuna)

• fix memory problem to increase batch size and speed

• gru

• gru-mlp (a modified version of gru)

• diff-add-cell (a designed RNN structure)

• ideas for improve fitting ??