Melissaratos L, Micheli-Tzanakou E
Department of Biomedical Engineering, Rutgers, State University of New Jersey, Piscataway 08855-0909.
J Med Syst. 1989 Oct;13(5):243-52. doi: 10.1007/BF00996458.
Optimization techniques have found many applications in science, engineering, and industry. In all applications, the best value of a "cost function" is sought in a well-defined domain; this cost function in general depends on many parameters. An iterative optimization technique has been developed (ALOPEX) that uses feedback in order to optimize the response of a system. The cost function for this process is problem dependent and therefore quite flexible. The method has been applied successfully to different optimization problems such as pattern recognition, receptive field studies in the visual system of animals, curve fitting, etc. We present two special purpose hardware implementations for ALOPEX. The first method takes time O(logN + logm) and uses O(mN2) processing elements. The second method takes O(logN + m) time and uses O(N2) processing elements. Our basic architecture is a binary tree with N2 leaves (equal to the length of the vectors) and therefore had depth O(logN). Different implications of the two approaches will be discussed including similarities with the biological visual process.
优化技术在科学、工程和工业领域有诸多应用。在所有应用中，要在一个定义明确的域中寻找“成本函数”的最优值；这个成本函数通常取决于许多参数。已经开发出一种迭代优化技术（ALOPEX），它利用反馈来优化系统的响应。此过程的成本函数取决于具体问题，因此非常灵活。该方法已成功应用于不同的优化问题，如模式识别、动物视觉系统中的感受野研究、曲线拟合等。我们展示了两种针对ALOPEX的专用硬件实现方式。第一种方法耗时O(logN + logm)，并使用O(mN2)个处理元件。第二种方法耗时O(logN + m)，并使用O(N2)个处理元件。我们的基本架构是一个具有N2个叶子节点（等于向量长度）的二叉树，因此深度为O(logN)。将讨论这两种方法的不同影响，包括与生物视觉过程的相似性。
