nn_theta_initializer_mlx: Internal: Calculate initial NN parameter values

Description

Calculate the initial NN parameter values, such that activation points are within the range between min_init and max_init defined in the un-converter Monolix model file

Usage

nn_theta_initializer_mlx(
  number,
  xmini,
  xmaxi,
  n_hidden = 5,
  theta_scale = 0.1,
  pre_fixef = NULL,
  time_nn = FALSE,
  act = "ReLU",
  beta = 20
)

Value

Vector of initial NN parameter values for one specific NN

Arguments

number: (string) Name of the NN, e.g., “1” for NN1(...)
xmini: (numeric) minimal activation point
xmaxi: (numeric) maximal activation point
n_hidden: (numeric) Number of neurons in the hidden layer, default value is 5
theta_scale: (numeric) Scale for input-hidden-weights initialization
pre_fixef: (named vector) Vector of pre-defined initial values
time_nn: (boolean) Definition whether NN is time-dependent (TRUE) or not (FALSE)
act: (string) Activation function used in the NN. Currently "ReLU" and "Softplus" available.
beta: (numeric) Beta value for the Softplus activation function, only applicable if act="Softplus"; Default to 20.

Author

Dominic Bräm

Details

theta_scale is the scale in which the weights from input to hidden layer are initialized, i.e., 0.1 initializes weights between -0.3 and 0.3; 0.01 initializes weights between -0.03 and 0.03
time_nn defines whether the NN is a time-dependent NN with the restriction that all weights from input to hidden layer are negative