The Complete SPRLIB & ANNLIB

learn_wavnet: - train a wavelet network on a dataset

SYNOPSIS

int learn_wavnet

(lf, l, proj, net, dset, cyc, opt, alpha, mom, L, U, c, d)

ARGUMENTS

`FILE`	*`lf`**	File to log errors to
`int`	`l`	Boolean indicating whether errors should be logged or not
`int`	`proj`	Specify if wavelets outside the domain of the function to be approximated should be truncated or not
`NET`	*`net`**	Wavelet network
`DATASET`	*`dset`**	Dataset used for training the network
`long`	`cyc`	Number of train cycles
`long`	`opt`	Flag (see below)
`double`	`alpha`	Learning rate in gradient descent wavelet network training
`double`	`mom`	Momentum term value
`double`	`L`	Lower value for scaling the dataset prior to training
`double`	`U`	Upper value for scaling the dataset prior to training
`int`	`c`	Parameter specifying the maximal amount of compression of a wavelon
`int`	`d`	Parameter specifying the tolerance for truncation of wavelons outside the domain of the function to be approximated

RETURNS

TRUE if an error occurred, FALSE otherwise.

FUNCTION

Training of a wavelet network using gradient descent minimization. For a reference in which this network is described, see the introductory text of this chapter. During training, the support of the wavelets in the hidden nodes may leave the domain of the function to be approximated or become too small (since dilation and translation parameters are adapted during training). Hence, hidden units may become useless. This can be tackled by truncating ("projecting") these parameters to minimum and maximum values, determined by the parameters compres_par and domain_par :

  
  compression_epsilon = (Upper - Lower)*CONST_XI/(num_wavelons*compres_par)
  eps_domain_lower = Lower - domain_par*(Upper - Lower);
  eps_domain_upper = Upper + domain_par*(Upper - Lower);

NOTE

The flag to be supplied as options is similar to the one that should be supplied with bp_adapt_net. However, no history of weights and values is stored and no batch-updating is allowed, hence options should be BPACCUM | BPUPDATE | BPOLDPRP. The function assumes that a wavelet network has been created and initialized.

The Complete SPRLIB & ANNLIB

`learn_wavnet`

- train a wavelet network on a dataset

SYNOPSIS

ARGUMENTS

RETURNS

FUNCTION

NOTE

SEE ALSO