A wawelet based heuristic to dimension Neural Networks for simple signal approximation

Before training a feed forward neural network, one needs to define its architecture. Even in simple feed-forward networks, the number of neurons of the hidden layer is a fundamental parameter, but it is not generally possible to compute its optimal value a priori. It is good practice to start from an initial number of neurons, then build, train and test several different networks with a similar hidden layer size, but this can be excessively expensive when the data to be learned are simple, while some real-time constraints have to be satisfied. This paper shows a heuristic method for dimensioning and initializing a network under such assumptions. The method has been tested on a project for waste water treatment monitoring.