Data mining-based adaptive regression for developing equilibrium speed-density relationships

In this paper, we developed a methodological framework to deal with traffic-stream modeling based on data mining, steepest-ascend algorithm, and genetic algorithm. The new method is adaptive in nature and has a greater flexibility and generality compared with existing methods. It provides an optimum overall fitting of the observed data. Specifically, the advantages of adaptive regression are that (1) knot positions and model parameters are estimated optimally and simultaneously using genetic algorithm, and presetting of knot positions can be performed in terms of either density or speed; (2) the method is automatic and data driven, and it will always find out the best fitting model to site-dependent actual traffic data; and (3) the user has a great flexibility to specify the degree-model continuity and to define and add new basis functions that are parsimonious and fit better into the traffic data in some regime of speed-density relation. The proposed method and developed computer software package MiningFlow will be beneficial to traffic operations and traffic simulation.