State Space Analysis and Design for Neural Networks

该项目的总体目标是开发一套定义良好的性能指标和评估技术，以取代目前在神经网络设计中使用的许多临时尝试。本课题研究了两种神经网络的状态空间：异步二元神经网络（ABNN）和自适应共振理论模型（ART1）。该项目专注于这两种结构对现实模式环境的适应，这些环境是短期的，但不是长期的。具体目标包括分析模型状态空间中的动态行为，以及引入结构和算法修改，以提高它们作为模式识别器的性能。将使用基于马尔可夫链理论、分支过程和最优停止规则的分析方法。该分析将通过在串行计算机上的模拟以及在模拟计算机上的真正并行实现来验证。

The overall objective of this project is to develop a set of well definedperformance indices and evaluation techniques to replace many ad-hoc cut-and-try efforts currently used in neural network design. This project is a studyof the state space of two neural networks: the asynchronous binary neural network (ABNN), and the Adaptive Resonance Theoretic model (ART1). The project concentrates on the adaptation of these two structures to realistic pattern environments, which are short-term but not long term stationary. Specific goals include the analysis of the dynamic behavior in the model's state space, and the introduction of structural and algorithmic modifications that will improve their performance as pattern recognizers. An analytical approach based on Markov chain theory, branching processes, and optimal stopping rules will be used. The analysis will be verified through simulations on a serial computer as well as a truly parallel implementation on an analog computer.