标题: | 空间分割最佳化方法与多层类神经网路学习法则及其应用 SPACE-DECOMPOSITION OPTIMIZATION METHODS AND LEARNING ALGORITHMS FOR MULTILAYER NEURAL NETWORKS WITH APPLICATIONS |
作者: | 刘锦松 Chin-Sung Liu 曾锦焕 Ching-Huan Tseng 机械工程学系 |
关键字: | 非限制条件最佳化;平行最佳化;空间分割最佳化;类神经网路训练法则;平行训练法则;Unconstrained optimization;Parallel optimization;Dpace decomposition optimization;Neural networks training algorithm;Parallel training algorithm |
公开日期: | 1998 |
摘要: | 最佳化方法已经广泛的应用在工程的设计问题上,因此本论文将发展一套新的空间分割最佳化方法。此方法可以使用在同步及非同步的平行处理器上,而能够加快最佳化方法执行的速度。空间分割最佳化方法亦可以处理多层线性、非线性及数位式类神经网路的训练问题。由于类神经网路本身具有平行计算及可适应性的功能,因此本论文亦利用空间分割最佳化方法,发展出一套较有效率的类神经网路学习法则,如此将可以减少类神经网路在训练上所需要的计算时间。而一理想的类神经网路电脑必须具备硬体及学习法则的可实现性,现今由于数位科技的进步,微处理机的性能已大幅度的改进,因此本论文亦将所发展出来的多层数位式类神经网路训练法则应用在一般数位式微控制器上,因此这种类神经网路电脑将具有较低成本及较高执行速度的特性,可以使用在一般的类神经网路应用上。 Optimization techniques are being used in a wide spectrum of industries to enhance the design of engineering systems. Therefore, this study investigates a set of new developed optimization techniques called space-decomposition optimization (SDO) algorithm. These new developed optimization techniques can be applied to synchronous and asynchronous parallel processors to enhance the optimization efficiency. They can be also applied to the training of multilayer neural networks, including linear, nonlinear, and binary neurons. Due to the adaptive nature, the neural networks offer a parallel-processing paradigm that could be user-friendlier than the conventional approaches. Although most applications demand very high throughput, most learning algorithms for neural networks are computationally intensive in nature. Therefore, efficient learning algorithms for neural networks are developed based on the SDO techniques to satisfy the demand in engineering. Digital technology has enjoyed a tremendous growth in CPU speed. An ideal digital neurocomputer provides an adaptive and flexible platform for neural network algorithms, and hardware implementation. Therefore, this study also develops efficient learning algorithms for multilayer binary neurons based on the SDO techniques. These algorithms can be easily applied to modern digital computers, such as microcontrollers. Therefore, the general-purpose digital neurocomputer can be implemented with a low-cost, high-speed, and flexible platform for neural applications. |
URI: | http://140.113.39.130/cdrfb3/record/nctu/#NT870489093 http://hdl.handle.net/11536/64773 |
显示于类别: | Thesis |