Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Larry E. Banta.
In this dissertation, a novel training algorithm for neural networks, named Parameter Incremental Learning (PIL), is proposed, developed, analyzed and numerically validated.;The main idea of the PIL algorithm is based on the essence of incremental supervised learning: that the learning algorithm, i.e., the update law of the network parameters, should not only adapt to the newly presented input-output training pattern, but also preserve the prior results. A general PIL algorithm for feedforward neural networks is accordingly derived, using the first-order approximation technique, with appropriate measures of the performance of preservation and adaptation. The PIL algorithms for the Multi-Layer Perceptron (MLP) are subsequently derived by applying the general PIL algorithm, augmented with the introduction of an extra fictitious input to the neuron. The critical point in obtaining an analytical solution of the PIL algorithm for the MLP is to apply the general PIL algorithm at the neuron level instead of the global network level. The PIL algorithm is basically a stochastic learning algorithm, or on-line learning algorithm, since it adapts the neural weights each time a new training pattern is presented. Extensive numerical study for the newly developed PIL algorithm for MLP is conducted, mainly by comparing the new algorithm with the standard (on-line) Back-Propagation (BP) algorithm. The benchmark problems included in the numerical study are function approximation, classification, dynamic system modeling and neural controller. To further evaluate the performance of the proposed PIL algorithm, comparison with another well-known simplified "high-order" algorithm, i.e., the Stochastic Diagonal Levenberg-Marquardt (SDLM) algorithm, is also conducted.;In all the numerical studies, the new algorithm is shown to be remarkably superior to the standard online BP learning algorithm and the SDLM algorithm in terms of (1) the convergence speed, (2) the chance to get rid of the plateau area, which is a frequently encountered problem in standard BP algorithm, and (3) the chance to find a better solution.;Unlike any other advanced or high-order learning algorithms, the PIL algorithm is computationally as simple as the standard on-line BP algorithm. It is also simple to use since, like the standard BP algorithm, only a single parameter, i.e., the learning rate, needs to be tuned. In fact, the PIL algorithm looks just like a "minor modification" of the standard on-line BP algorithm, so it can be applied to any situations where the standard on-line BP algorithm is applicable. It can also replace the standard on-line BP algorithm already in use to get better performance, even without re-tuning of the learning rate.;The PIL algorithm is shown to have the potential to replace the standard BP algorithm and is expected to become yet another standard stochastic (or on-line) learning algorithm for MLP due to its distinguished features.
Wan, Sheng, "Parameter incremental learning algorithm for neural networks" (2005). Graduate Theses, Dissertations, and Problem Reports. 2252.