Transcript Back-Propagation MLP Neural Network Optimizer

Back-Propagation MLP
Neural Network Optimizer
ECE 539
Andrew Beckwith
Back-Propagation MLP
Network Optimizer
Purpose
 Methods
 Features

Purpose




Configuring a Neural Network and its parameters
is often a long and experimental process with
much guess work.
Let the computer do it for you.
Design and implement a program that can test
multiple network configurations with easy setup.
Allow user to modify data properly by enhancing
important features and minimizing features with
little importance or detrimental qualities.
Methods





Use back-propagation algorithm with momentum
To test multiple configurations, use brute force method
and keep track of most successful configuration.
Only parameter user cannot control is the number of
neurons per hidden layer.
Each configuration is tested with 2, 3, 5, and 10 neurons
per hidden layer. The last test is a random initialization
between 1 and 10 for each layer.
Use hyperbolic tangent activation function for hidden
neurons and sigmoidal activation function for output
neurons. One could change this in the source code if
desired.
Features




Allow user to open data file, view mean and
standard deviation for each feature of each class
for modification purposes.
Allow user to enter ranges and number of trials
for parameters such as: max epoch, epoch size,
learning rate, momentum constant, and the
number of hidden layers.
Allow user to set a tolerance to achieve
maximum classification rate.
Allow user to view entire network – network
configuration, weight values, etc.