APPLYING NEURAL NETWORK MODELS TO THE CONSTRUCTION OF PRODUCTION RULES EXPERT SYSTEMS

Abstract

The development of urbanization, the dramatic increase in the population of large cities and, as a consequence, the increased tempo of life, have had a direct impact on the number of vehicles. That is why intelligent methods of traffic flow control to optimize traffic capacity are growing in popularity. The application of these methods makes it possible to use existing roads and highways with high efficiency without the need to build additional lanes, interchanges, ring roads and so on, which in particular is postulated by the strategy of the National Technological Initiative of the Russian Federation. The purpose of this study is to develop an algorithm, the idea of which is to form productive rules for the traffic flow control by directing the impact on its parameters. In this way, experts can assess the impact of certain characteristics on those that can be manipulated. The construction of the algorithm is done by forming neural network models followed by using the method of the most dominant rule and Garson's algorithm. Since the use of the algorithm involves expert group evaluation, it must have a logical (verbal) rule output in the following form, e.g: "if the slope is minimum, the width of the roadway is maximum, ... then the percentage of heavy vehicles is minimum". This transformation is carried out through the use of the membership function, which allows to fully describe the degree to which a certain parameter belongs to a certain fuzzy subset. In the problems of traffic flow control, it is recommended to use a triangular-shaped membership function. The study contains the results of computational experiments to determine the optimal partitioning of input parameters into fuzzy values to generate satisfying real-world conditions rules. When constructing the algorithm, special attention should be paid to assessing the quality of the neural network model. For this purpose such methods of estimation as root mean square error (RMSE) and logistic error function (LogLoss) are used. Data from loop and radar detectors describing the capacity in long-term operating areas on sections of transportation corridors were used as the initial data set for the numerical study. Sensitivity analysis based on the application of the finite increment formula was used to determine the most significant traffic flow parameters.