NEWS

hann 1.1-3

The new function combine makes predictions combining several neural networks possibly optimized with different classes.
The new function tune.hann helps to choose the hyperparameters (number of hidden neurons and slope of the activation function).
The new function hann is a wrapper for all types of neural nets. The new class "hann" has 'secondary' classes "hann1" or "hann3". There are several new methods: summary(), plot(), coef(), fitted(), str(), and labels().
The optimized networks now include an element $fitted with the final signals, and an element $labels with the labels of the input classes.
The new function binarize helps to prepare the binary patterns from vectors or matrices of numeric values (e.g., read from image files).

A bug has been fixed in the calculations of gradients of hann3() so convergence should be improved (and faster).
The internal code used to allocate new memory at each iteration: this is fixed.

The internal code of hann3() has been optimised so it should be at least twice faster.