Meta-cognitive learning for pattern classification

The goal of this project is to develop an optimized learning algorithm for structured, unstructured and random knowledge presented from a stream of data set. The current algorithm available provides a certain amount of accuracy but it is not optimized to...

Full description

Saved in:
Bibliographic Details
Main Author: Chng, Adrian Yong Hao
Other Authors: School of Computer Engineering
Format: Final Year Project
Language:English
Published: 2014
Subjects:
Online Access:http://hdl.handle.net/10356/59018
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:The goal of this project is to develop an optimized learning algorithm for structured, unstructured and random knowledge presented from a stream of data set. The current algorithm available provides a certain amount of accuracy but it is not optimized to perform better. The purpose of this project is to make use of this algorithm and optimized it further such that the machine will be able to learn automatically and classify the input data into the correct classes where the data belong after learning through some training data. This will then result in a better accuracy of the algorithm. The project was carried out in two phrases. In the first phrase, I was tasked to read up on radial basis function neural networks, projection based learning and how machines can be used to solve classification problems. I was then given some sample data and the algorithm code to read up, analyze the flow of the program and test the data. In the next phrase, I was tasked to modify the algorithm code to integrate Projection Based Learning (PBL) with Particle Swarm Optimization (PSO). PSO is used to optimize the parameters of the algorithm using cross-validation such that it can be applicable to the different training samples to generate a more accurate result. In total, 9 data sets were used in this project. This report provides detailed information on how the algorithm were designed and implemented to suit the needs of this project. ii The results found were that by adjusting the four parameters: skip threshold, initial adding error threshold, initial learning error threshold and limit for reserve samples, better accuracies can be attained. These parameters also vary differently to different input data set and they will be the focus for adjusting the algorithm. In conclusion, algorithms are able to classify samples accurately into the different classes to certain degree of accuracy.