An Evolutionary Credit Apportionment Algorithm for Neuromolecular Computer Design

神经分子计算机设计的进化信用分配算法

• 批准号: 9109860
• 负责人: Michael Conrad
• 金额: $ 20.1万
• 依托单位: Wayne State University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 1991
• 资助国家: 美国
• 起止时间: 1991-09-01 至 1995-02-28
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9109860&HistoricalAwards=false
• 关键词: Evolutionary; Credit; Apportionment; Algorithm; Neuromolecular

The goal of this research is to apply a newly developed evolutionary learning algorithm to neuromolecular computer design. The distinguishing feature of neuromolecular networks is that the input-output behavior of the neuron-like modules is controlled by significant intraneuronal dynamics, including macromolecular recognition mechanics and transduction of input signals to a form that can be recognized by marcomolecules. The working hypothesis is that significant biological information processes are mediated by systems of this type; moreover, the rapid development of molecular electronics technologies should in the future allow for bona fide molecular implementations. A simulation system that can be used for designing such systems and that also has artificial intelligence applicability in the context of existing silicon technologies has been developed. During the past year an advanced evolutionary learning algorithm that utilizes a memory manipulation capability to assign credit to individual neurons in a network has been also been developed. Credit apportionment is based on input control and reward sharing in a population of virtual networks. The algorithm has been tuned on connectionist neural networks and has been tested successfully on neuromolecular networks with relatively simple intraneuronal dynamics. Cellular automaton dynamics will be used to represent signal processing in the cytoskeleton, and the credit apportionment algorithm will be used to train networks of "cytoskeletal neurons". A key objective is to build " evolution friendliness" into the intraneuronal and network dynamics on which evolutionary algorithms act by simulating the type of structure-function relations that enable natural biological systems to learn efficiently through evolution. High dimensionality is critical since it increases the likelihood of finding easily traversible pathways between functionally useful systems. The study will focus on the effects of increasing the intraneuronal cellular automaton dynamics in neuromolecular networks and memory- manipulation algorithms capable of orchestrating special purpose cytoskeletal neurons into coherent groups capable of performing complex pattern recognition/effector control tasks.

本研究的目的是将一种新开发的进化学习算法应用于神经分子计算机设计。神经分子网络的显著特征是神经元样模块的输入输出行为是由重要的神经元内动力学控制的，包括大分子识别机制和输入信号转导到一种可以被小分子识别的形式。工作假设是，重要的生物信息过程是由这种类型的系统介导的；此外，分子电子技术的快速发展应该在未来允许真正的分子实现。一个仿真系统，可用于设计这样的系统，也具有人工智能的适用性，在现有的硅技术的背景下已经开发。在过去的一年中，一种先进的进化学习算法也被开发出来，该算法利用记忆操纵能力将信用分配给网络中的单个神经元。在虚拟网络群体中，信用分配基于输入控制和奖励共享。该算法已在连接主义神经网络上进行了调整，并已在具有相对简单的神经元内动力学的神经分子网络上成功地进行了测试。将使用元胞自动机动力学来表示细胞骨架中的信号处理，并使用信用分配算法来训练“细胞骨架神经元”网络。一个关键目标是将“进化友好”构建到神经元内和网络动力学中，进化算法通过模拟结构-功能关系的类型来发挥作用，从而使自然生物通过进化有效学习的系统。高维是至关重要的，因为它增加了在功能有用的系统之间找到容易穿越的路径的可能性。本研究将集中于增加神经分子网络中神经元内细胞自动机动力学的影响，以及能够协调特殊目的的记忆操作算法细胞骨架神经元组成连贯的组，能够执行复杂的模式识别/效应控制任务。