Partially-Connected Artificial Neural Networks for Time Series

This program can generate, train, and test any series of partially-connected artificial neural networks against a time series function, with the only restriction being that the networks have three layers.

Introduction

This program is based on the back-propagating neural network implementation of Neil Schemenauer which has been placed in the public domain. You can find an original copy of his work here.

The primary use of this program is not to train and test a single neural network, but to test all permutations of a given node configuration. Weights are connected and disconnected until all possible networks have been trained and tested, with performance indicators being logged and returned as a CSV file.

Requirements

To use pcann-time-series.py, the following utilities are required:

• Python 2.6 or greater.

To analyze the data generated by pcann-time-series.py, you must have the ability to work with CSV output.

Usage

pcann-time-series.py - Train and test all potential networks.
                       Outputs CSV dataset to STDOUT.
                       Networks must be 3 layers.
Usage: pcann-time-series.py [OPTIONS]
-h|--help                  Display this help
-i|--num-input num         Set the number of input nodes
-d|--num-hidden num        Set the number of hidden nodes per layer
-o|--num-output num        Set the number of output nodes
-f|--function sin|sinh|saw The time series function to use
-w|--freq num              The frequency of the wave
-a|--amp num               The amplitude of the wave
-p|--phase num             The phase of the wave
--train0 theta             The first training input
--trainf theta             The last training input
--test0 theta              The first test input
--testf theta              The last test input
--threads num              The number of threads to use