Date of Award
Doctor of Philosophy
Electrical and Computer Engineering
This research proposes a novel Adaptive Competitive Self-organizing model, shortly named ACS, with applicability for real-time clustering and vector quantization. The model is designed based on sets of Ordinary Differential Equations (ODE’s) and free of any external controls. These properties make it suitable for hardware implementation and real-time applications. This classifier considers as unsupervised Neural Network (NN) since it doesn’t have any prior knowledge on input pattern’s classes. The design of this classifier is based on developing an energy function, constructed based on the sum of Lorentzian functions, where they correspond to the clusters of similar input patterns. The defined energy function is a form of Lyapunov function, and it guarantees trajectories of weights, which are labeling a set of similar input pattern will finalize in a set of isolated equilibrium points. The stability of those equilibrium points is investigated. Valleys on the surface are the representation of clusters, and they are defined with their parameters such as depth, width and vigilance parameter. These control parameters are effective on convergence speed, the accuracy of labeling, etc. We are going to study the level of effectiveness of those parameters on clustering assignments with different illustrations. All these parameters need to be dynamically adjusted in the model, resulting in the highest level of self-adjustment. To comprehend the way clustering occurs in ACS, two main processes are utilized to pattern clustering: learning and recalling. In learning phase the surface of energy modifies as sets of weights are self-adjusting themselves in a competitive manner to label/cluster an exposed input pattern on the surface of energy function, while recalling phase is as a newly exposed input pattern accommodate itself into existence similar cluster, which has the shortest Euclidean distance from it. The effectiveness of ACS model is demonstrated with implementing it on both real and artificial data sets as well as comparing with other well-known clustering methods. ACS method showed a better clustering performance in some categories and an overall comparable rendition. System dynamics is simulated with two optimizers Gradient Descent (GD) and Adaptive Momentum Gradient Descent (AMGD), in cooperation with a competition mechanism based on the Lotka-Volterra competition exclusion. Simulation results indicate the effectiveness of Adaptive Momentum Gradient Descent (AMGD) in achieving the optimal convergence speed of ACS in doing clustering assignments, in compare with GD and classical Momentum method.
This dissertation is only available for download to the SIUC community. Current SIUC affiliates may also access this paper off campus by searching Dissertations & Theses @ Southern Illinois University Carbondale from ProQuest. Others should contact the interlibrary loan department of your local library or contact ProQuest's Dissertation Express service.