Research And Application Of Model Robustness For Traffic Prediction And Decision

With the continuous development of technologies such as big data collection and global satellite positioning systems,transportation systems are becoming increasingly complex and busy,the importance and safety of intelligent transportation are growing trend.Research in the field of traffic prediction based on deep learning has been an important driving force for the development of intelligent transportation systems in recent years.Various traffic models in the field rely on accurate traffic data.However,traffic data has the characteristics of explosive growth in quantity,uneven quality and strong concealment of errors,phenomena such as performance degradation and classification errors of traffic models have often occurred,the entire intelligent transportation system has easily overlooked distortion issues,and the robustness of traffic models faces serious challenges.Aiming at the above-mentioned phenomenon,this project focuses on the research and application of model robustness for traffic prediction and decision-making,as follows:(1)Conducted in-depth and extensive survey on government open transportation datasets and transportation datasets commonly used by academia.Applying traditional data quality evaluation systems to traffic data can lightweight use the relative closeness D between the data and the ideal solution to evaluate the quality of traffic data sets.(2)Tapping the idea of predictive uncertainty theory,a research system for traffic data driven model robustness is established.Two robustness evaluation indicators are proposed: information conflict value indicator and information invalidity indicator.These two indicators are used to measure the robustness of traffic models,and the design of a traffic model robustness evaluation framework is given.(3)Taking advantage of the research method of traffic model robustness based on prediction uncertainty,we conducted a fair,reasonable,multi-directional and extensive robustness evaluation of the four most advanced traffic prediction models and one classic traffic decision-making model in the current scenario,making up for the defect that traffic models cannot quantify their uncertainty,and understanding the robustness related performance of various models.(4)A new method for generating test data based on conflict values with strong authenticity has been proposed,which can generate the adversarial samples required by researchers.A comparative experiment is designed,and the experimental results show that the Conflict method is superior to the Sim GAN based method in generating test data,which proves the effectiveness and progressiveness of this method.