multiple hidden layers problem

[Alejandro-FQA]

When using the NQS and more than one hidden layer, NaN appear already during imaginary time propagation.

Possible checks:

• Integration step
• Regularization of QGT
• Inversion method of QGT

[Alejandro-FQA]

Currently unable to reproduce the issue with NaNs, but the wavefunction collapses to '0' as the parameters increase 1 order of magnitude suddenly.

Case:

hidden_layers = [4,4]
pm.dt = 0.1

[Alejandro-FQA]

Major issues identified:

lambda_reg $\in\mathbb{R}$ works better for imaginary time propagation.
lambda_reg $\in\mathbb{C}$ works better for real time propagation.
lambda_reg = 1e-3 seems to be the optimal value.

Cases with 1, 2 and 3 layers have been checked.

Each additional layer increases the computation time a 50% (w.r.t. single layer NN).
Same number of parameters in multiple layers perform slightly better.

For the cases analyzed, NaNs appeared only for 3 layers, when keeping dt = 0.1.

Smaller dt might avoid further NaNs.

Case example 1-2-2-4-1 shown below with e_error = 1e-4.
Simulation stopped after parameters suddenly increase.