1. UKM Learning and Research Repository
2. Master Thesis / Tesis Master
3. Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Title:	Enhanced time delay neural network for temporal pattern recognition
Authors:	Anahita Ghazvini (P72674)
Supervisor:	Siti Norul Huda Sheikh Abdullah, Prof. Madya Dr.
Keywords:	Temporal pattern recognition Crime data Real-time data Dissertations, Academic -- Malaysia
Issue Date:	21-Jan-2016
Description:	Temporal pattern recognition is a recurrent architecture with the purpose of real-time data prediction such as economic forecasting or systems modeling. In conjunction to time delay, an associated dynamic classifier system is commonly predicted based on nonlinear methods such as Time Delay Neural Network (TDNN) despite of its instability and poor performances. Example of TDNN classifier systems are nonlinear autoregressive with exogenous input (NARX) and back propagation through time (BPTT). In general, TDNN is a fully recurrent network which conquered the weakness of static networks but it still has limitation in identifying the dynamical systems using single objective functions. Thus, this single objective function can be further improved by minimizing TDNN error directly during training phase. Safe City Monitoring System (SCMS) is a crime mapping tool that helps the Royal Malaysia Police (RMP) to record temporal report data digitally namely Police Reporting System (PRS). It is able to analyze crime data, produces crime hotspot points and crime frequency, and identifies potential crime locations and time without any automatic crime spatial analysis. Hence, PDRM predicts the next crime time, location and suspect’s age solely based on evidences such as report, past experiences, intuitive or rule of thumb information from witnesses which is time and cost expensive. The objectives of this thesis are (1) to propose an improved NARX model based on combination objective functions, and (2) to develop a commercial serial crime spatial analysis based on the improved NARX model. After receiving summation of multiplication input-weight and bias value, the improved NARX model was proposed by summing two objective function values from hyperbolic tangent (Tan Sigmoid) and Radial Basis Function (RBF) in a single hidden layer with 10 nodes before estimating the final output value of the time delay neural network. In this research, three benchmark temporal datasets namely, Dow Jones Index, Monthly River flow in cubic meters per second, and Daily temperature in degree centigrade were examined. Two UKM-PDRM datasets namely, “Suspect & Capture” and “Crime Plotting” datasets have been obtained with ethical approval from PDRM and SCMS respectively. The improved NARX was evaluated and compared with the state of art methods NARX and Back Propagation Through Time (BPTT). The average accuracy of the Dow Jones Index, Monthly River Flow in Cubic Meters per Second (cms), and Daily Temperature in Degree Centigrade (°C) datasets for the reliability of the improved NARX, NARX, and BPTT model are about 82.15%, 74.21%, and 49.71% respectively with significant p-value of 3.75E-25 (p<0.05) whereas the commercial crime series achieved about 80.76%, 76.03%, and 37.82% respectively with significant p-value of 0.0125 (p<0.05).,Certification of Master's/Doctoral Thesis" is not available
Pages:	181
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.