Code Description

Click to jump to a particular section of this page.

Code Reference

      The original version, documented version and ported versions of the code used can be found through the links below (or attached sheets if hardcopy).

• Original code from book
• Documented code from book
• Ported code

      The file and data handling code used in this project is from the Weka project (www.cs.waikato.ac.nz/ml/weka/).

      Weka is a collection of machine learning algorithms for solving real-world data mining problems. It is written in Java and runs on almost any platform. The algorithms can either be applied directly to a dataset or called from your own Java code. Weka is also well-suited for developing new machine learning schemes. Weka is open source software issued under the GNU public license.

      The only Weka code used in this project was that which handles files and various data manipulation tasks. It was also used to collect statistics on the performance of the Back Propagation algorithm implemented.

The Back Propagation algorithm implementation provided by the book was also used in this project.

Back to Top

      The code from the book was ported to Java. There were several reasons for this. The C code supplied handled input that was in the form of an image, hence, containing only numeric data to be handled. The file parsing and data handling functionality needed to handle the Census-Income data set was done in Project 2 for Java. This presented several options:

1. Port file and data handling functionality to C
2. Call native C functions from Java
3. Port Neural Network - Back Propagation C code to Java

Functions and sections used directly from backprop.c

• double squash ( double x )
• bpnn_ranodmize_weights ( double **w, int m, int n )
• bpnn_zero_weights ( double **w, int m, int n )
• bpnn_internal_create ( int n_in, int n_hidden, int n_out )
  • Note: Just array declaration - allocation was used. The ported code determines the number of input and output nodes needed for the specified data set. This can be found in the BackPropagation.buildClassifier method.
• bpnn_create ( int n_in, int n_hidden, int n_out )
  • See note above for bpnn_internal_create
  • Further more, the option to initialize weights with zero values was not provided in the ported version.
• bpnn_layerforward ( ... )
• bpnn_ouput_error ( ... )
  • Note: since this function is always called with the same arguments which are available in the ported class BackPropagation , the argument list was removed.
  • Note: The err value became the return value for the method in the ported version
• bpnn_hidden_error ( ... )
  • Please see comments for bpnn_output_error
• bpnn_adjust_weights ( ... )
  • Note: eta and momentum arguments were removed from argument list and are contained and maintained by the BackPropagation object. These can be specified with command line options in the ported version.
• bpnn_feedforward ( BPNN *net )
  • Note: argument removed since the method exist inside the object equivalent of BPNN
• bpnn_train ( ... )
  • Note: arguments were removed from ported version. The sum of the error values became the return value of the ported method. This was done since the calling function always added these two values together. A more elegant solution is possible with more time.
  • Note: where bpnn_train explicitly called bpnn_layerforward twice, the ported version calls the Feed Forward method. This is equivalent, code wise, of how bpnn_train uses Layer Forward.

Back to Top