Applicability of Neural Network Models to the Future Traffic Management Systems.

本文主要研究了基于神经网络模型的自组织交通控制系统的开发。为了实现自组织交通管理系统，引入了神经网络模型。首先，将包含训练过程和优化过程的多层神经网络模型应用于单个交叉口的最优交通控制问题；结果表明，在训练过程中，采用反向传播方法对训练操作进行几次迭代，就能建立稳定的输入输出关系。在优化过程中，采用柯西法与反馈法相结合的逐步优化方法，可以得到较好的控制变量逼近序列。与单独使用反馈方法的结果相比，该方法能够避免陷入局部最小值，并迅速达到全局最小值。其次，针对由多个交叉路口组成的道路，神经网络的突触权值快速增加的问题，提出了一种只与一个交叉路口相关的神经元相互连接的多重分割模型。这种多重分割模型不仅提高了计算时间，而且提高了对未训练模式的估计精度。第三，为了解决偏移量的优化和交通情况的变化，提出了一个具有三个输入源的复杂神经网络，一个输入源用于分割，一个输入源用于偏移量，另一个输入源用于流入交通量。根据输入源的神经网络连接方式，将该模型分为四种类型。最后，将互联神经网络模型应用于动态交通管理系统中与路线引导系统密切相关的一般最小成本流问题。为了得到最优解，采用Hopfield模型。即，基于总旅行时间对应于神经网络系统总能量的假设，给出了神经系统的突触权值和输入偏置。考虑了各路段流量不超过路段容量的路段流量约束，以及平均行驶时间取决于路段流量。结果表明，通过指定适当的目标函数权重因子，Hopfield算法能够给出与解析解较为一致的近似解。少

This study is mainly concerned with developing of a self-organizing traffic control system using neural network models. To realize a self-organizing traffic management system, neural network models were introduced. First, a multilayered neural network model, consisting of both a training process and an optimization process, was applied to an optimal traffic control problem of a single intersection. It was shown that in the training process a few iterations of the training operation by the backpropagation method was able to build up a steady input-output relationship. In the optimization process a stepwise method that combined the Cauchy method with a feedback method was able to produce a good approximated sequence of control variables. Compared to results by the feedback method alone, this stepwise method was able to avoid entrapment into local minimums and reach a global minimum promptly.Second, to deal with the rapid increase of synaptic weights of neural networks for roads that cons … More ist of several intersections, a multiple split model, in which only neurons that were related to an intersection were connected to each other, was proposed. This multiple split model improve d not only the computation time but also the estimation precision for untrained patterns.Third, to deal with the optimization of offsetsand the variation of traffic situations, a complicated neural network, which has three input sources, one for splits, one for offsets and the other for inflow traffic volumes, was proposed. This model was classified into four types by how the neural networks for those input sources were connected.Finally an interconnected neural network model was applied to a general minimum cost flow problem that was closely associated with a route guidance system in dynamic traffic management systems. To obtain the optimal solution the Hopfield model was used. That is, synaptic weights and input biases of the neural system were formulated based on the assumption that the total traveling time corresponds to the total energy of the neural network system. Considered were constraints on link flow where each link flow does not exceed the link capacity, and where the average traveling time depends on link flow. It was found that by specifying proper weight factors of the objective function, the Hopfield algorithm was able to give approximated solutions that were in good agreement with analytical ones. Less