Advanced Traffic Control System on Urban Road Network by Neural Network Models

利用人工智能技术，我们开发了一种逐步优化城市街道信号时序参数的方法，如分割和偏移。该方法分为两个过程，一个是训练过程，一个是优化过程。在训练过程中，我们使用了两个神经网络模型；多层模型和Kohonen特征映射模型。前一种模型建立了信号时序参数与目标变量之间的输入输出关系。后一种模型提高了计算效率和估计精度。在优化过程中，为了避免陷入局部最小值，我们使用了两种人工智能方法；柯西机器和遗传算法。我们调整了定时参数，使延迟时间和停止频率的总加权和最小。我们将两种人工智能方法的解决方案与传统方法的解决方案进行了比较，并证实它们对于未来建立先进的交通控制系统是有用的。在此基础上，结合kohonen Feature Map技术建立了多层神经网络模型，描述了交通密度、交通流率和空间平均速度等交通变量之间的宏观关系。与解析回归方法的比较表明，神经网络方法大大提高了回归系数，并能很好地描述变量间的非线性和不连续行为。这种自组织关系有助于精确地模拟交通流量，并有效地检测事故。

Using artificial intelligence techniques, we developed a stepwise method to optimize signal timing parameters, such as splits and offsets, on an urban street. The method is separated into two processes, a training process and an optimization process. In the training process, we used two neural network models ; a multilayr model and Kohonen Feature Map model. The former model builds an input-output relationship between the signal timing parameters and the objective variable. The latter model improves the computational efficiency and the estimation precision. In the optimization process, to avoid the entrapment into a local minimum, we used two artificial intelligence methods ; the Cauchy machine and a genetic algorithm. We adjusted the timing parameters so as to minimize the total weighted sum of delay time and stop frequencies. We compared the solutions by both artificial intelligence methods with those by a conventional method and confirmed that they were useful for establishing advanced traffic control systems in the future.Next we described the macroscopic relationships among the traffic variables such as density, traffic flow rate, and space mean speed by a multilayr neural network model which was combined by the kohonen Feature Map technique. Comparison with analytical regression method proved that the neural network approach improves the regression coefficient a great deal and describe well the non-linear and discontinuous behavior among those variables. Such self-organizing relationships serve to simulate the traffic flow precisely and to detect incidents efficiently.