USP Theses Collection


Menu



A B C D E F G H I J-L M N O P Q-R S T U-V W-Y
close this section of the library Neural networks (Computer science)


View the PDF document Multi-objective cooperative neuro-evolution for chaotic time series prediction
Author: Chand, Shelvin
Institution: University of the South Pacific.
Award: M.Sc.
Subject: Neural networks (Computer science), Time-series analysis
Date: 2014
Call No.: Pac QA 76 .87 .C53 2014
BRN: 1198119
Copyright:Under 10% of this thesis may be copied without the authors written permission

Abstract: The use of neural networks for time series prediction has been an important focus of recent research. Multi-objective optimization techniques have been used for training neural networks for time series prediction in the past. Cooperative coevolution is an evolutionary computation method that decomposes the problem into subcomponents and has shown promising results for training neural networks. This thesis presents a multi-objective cooperative coevolutionary method for training neural networks for time series prediction where the time series data sets are preprocessed to obtain the dierent objectives. Data sets with dierent time lags are used as the objectives to be optimized. The method is tested on benchmark data sets including both real world and simulated time series problems. The results show that the multi-objective approach is able to improve the overall prediction accuracy while using one generalized neural network for predicting data sets representing dierent time-lags. A prototype of a mobile application for nancial prediction is also given for potential investors to use on their Android based mobile devices.
View the PDF document Governing robotic motion via a single-layer artificial neural network
Author: Prasad, Avinesh
Institution: University of the South Pacific.
Award: Ph.D.
Subject: Robots -- Motion -- Mathematical models , Neural networks (Computer science)
Date: 2012
Call No.: Pac TJ 211 .4 .P73 2012
BRN: 1190651
Copyright:10-20% of this thesis may be copied without the authors written permission

Abstract: This thesis addresses the design and implementationof a new approach to address the motion planning and control problem of two-dimensional and three-dimensional robots. The approach of solving the problem is decoupled. Firstly, there is an establishment of a uniquely tailored velocity algorithm which is capable of driving the robot from its initial position to the target positionand remain there forever. Secondly, a supervised single-layer artificial neural network, which employs an arctan activation function, is used to model the turning/steering angle of the robot ensuring that the robot steers safely pass an obstacle. A simple and easy method is also developed for obtaining a set of data for training the network. The training data is obtained using computer simulationwhere the initial path is traced by the user. With data obtained, the neural network is then trained using the least square method. In this thesis, the purpose of the single-layer artificial neural network is to control the motion of the robot when the robot is in the sensing zone of the obstacle. Otherwise, the velocity algorithm is enough to drive the robot to its target in the absence of obstacles. Various kind of obstacles such as fixed, moving and artificial obstacles are studied within the collision avoidance scheme. It has been noticed that our method is efficiently useful in designing control laws that can incorporate these obstacles. The mechanical singularities of the robot are carefully taken into account either by treating it as an artificial obstacle or by incorporating it into the control laws. Moreover, the stability of the system is also studied via the Direct Method of Lyapunov. The work carried out naturally falls into three distinctive parts.
Disclaimer & Copyright l Contact Us l
© Copyright 1968 - 2018. All Rights Reserved.
USP Library
The University of the South Pacific
Laucala Campus, Suva, Fiji
Tel: +679 323 1000