Ee4305 Lab 2

EE 4305 Introduction to Fuzzy/Neural Systems

Classification of Cardboard Papers Using a Multilayer Perceptron

Name: Low Kang Jiang Matric Number: A0074752B Email Address: a0074752@nus.edu.sg Supervisor: Associate Prof. Chen Chao Yu, Peter

0|Page

1. Objectives

 Competence in using a numerical computational software(Matlab) to develop a program for implementing a Multilayer Perceptron(MLP) incorporated with the errorbackpropagation algorithm Understand of the principles and issues in training and testing a MLP



2. Project Description

In the project, a MLP is to be trained to classify five types of cardboard papers based on their odors. Detecting odors in cardboard papers is an important issue in the packaging industry. It is an essential requirement that the material used for packaging does not introduce unwanted odors into the product(eg. foodstuffs) contained in the package. The odor of cardboard paper is measured by a hybrid gas array sensor(a.k.a an “electronic nose”). The odors from five different types of cardboard paper from commercial manufacturers were recorded with the elctronic nose. The MLP will be trained to determine the type of an unknown sample based on the data measured by the electronic nose.

3. Training, Testing and Evaluation

10 Hidden Layers

Activation Function for Hidden layer Hyperbolic Tangent Sigmoid Training Function Levenberg-Marquardt Backpropogation Biasing Condition 0 Initial Weight [-0.5 0.5] Weight Learning Function Gradient Descent with Momentum Performance Function Sum Square Error Hidden Layer Input Number 10 Hidden Layer Input Size 15 Maximum Epochs 1000 Stop Criterian SSE < 0.00001(1e-6) Learning Rate 0.001 Epochs Step 1 Hit Defining Square Error 0.001 Table 1: Parameter Setting for 10 Hidden Layers Epochs of Training 24 Training Sum Square Error(Best) 9.78e-8 Training Hit Number 288 Training Miss Number 0 Testing Sum Square Error 2.7e-3 Testing Hit Number 288 Testing Miss Number 0 Maximum Percentage of...