Marker-Propagation Networks: A Quantitative Analysis

标记传播网络：定量分析

• 批准号: 9406998
• 负责人: Dan Moldovan
• 金额: $ 23.5万
• 依托单位: Southern Methodist University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 1994
• 资助国家: 美国
• 起止时间: 1994-12-01 至 1998-11-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9406998&HistoricalAwards=false
• 关键词: Marker; Propagation; Networks; Quantitative; Analysis

Marker-propagation networks (MPNs) represent a powerful computational model especially suitable for artificial intelligence and other applications where information is often uncertain and incompletely specified. MPNs refer to networks of active nodes and labeled links. The power and novelty of this model derives from the fact that the processing instruments are programmable data patterns of flexible length called markers. Node processing is activated by the arrival of markers. Marker propagations through the network are not driven by destination addresses, instead, they depend on local conditions in the nodes, network topology and the information programmed in each marker. This is a major difference between MPNs and other parallel processing models, and it is the reason why MPNs are capable of handling weakly or incompletely specified applications. Previous work in the area of marker--propagation processing assumed much simpler models. The model studied in this project operates asynchronously, and is highly programmable. Some of the critical issues related to the implementation of MPNs on existing parallel computers are possible programming paradigms, sources of parallelism, load balancing, design tradeoffs, and others. The approach taken in this project is a quantitative one, meaning that the model is implemented on several parallel computers, performance is measured and the results compared. The MPN model offers scalability and has the potential of being a preferred architecture for future supercomputers dedicated to knowledge processing applications.

标记传播网络（MPN）代表了一种强大的计算模型，特别适合人工智能和其他信息通常不确定且不完全指定的应用。 MPN 指的是活动节点和标记链路的网络。 该模型的强大功能和新颖性源于这样一个事实：处理仪器是长度灵活的可编程数据模式，称为标记。 节点处理由标记的到达激活。 通过网络的标记传播不是由目标地址驱动的，相反，它们取决于节点中的本地条件、网络拓扑和每个标记中编程的信息。 这是 MPN 与其他并行处理模型的主要区别，也是 MPN 能够处理弱指定或不完全指定的应用程序的原因。 标记传播处理领域的先前工作假设了更简单的模型。 本项目研究的模型异步运行，并且具有高度可编程性。 与在现有并行计算机上实现 MPN 相关的一些关键问题包括可能的编程范例、并行源、负载平衡、设计权衡等。 该项目采用的方法是定量方法，这意味着模型在多台并行计算机上实现，测量性能并对结果进行比较。 MPN 模型提供了可扩展性，并且有潜力成为未来致力于知识处理应用的超级计算机的首选架构。