Identification and Adjustment of Implausible Outputs of Machine Learning Models for Enhancing the Reliability of Switching Time Predictions for Traffic-Actuated Signals

Estimates of future switching times for traffic signals are mandatory information to enable Green Light Optimized Speed Advisory (GLOSA) or similar eco-drive applications. Because of the adjustment of the signalization to match current traffic situations, obtaining reliable switching time estimates for traffic-actuated signals is challenging because they are not derivable based on bare statistics. An expectation in machine learning (ML) algorithms is that they can identify and learn existing causalities and relations based on provided data. Considering this, the most recent research focusing on predicting traffic-actuated signals’ switching times utilizes predictive ML models trained with historical switching times and detector data. Nevertheless, these algorithms probably cannot comprehensively capture each intrinsic relation and boundary condition existing in signal programs. Therefore, a general exclusion of implausible predictions, which, for example, violate permitted green windows or durations, should not be taken for granted. Considering this, this paper presents a methodology for testing the plausibility of predictive ML model outputs, enabling further evaluations of the reliability of applied predictive models. The methodology further incorporates procedures for preventing, identifying, and adjusting implausible switching time predictions, aiming to enhance the reliability of switching time predictions for traffic-actuated signals. The performance of the developed procedures is evaluated using prediction experiments conducted with historical data from multiple real-world traffic-actuated signal systems. The results indicate that the developed procedures can contribute to enhanced forecast accuracy. The magnitude of the achievable accuracy gains, thereby, depends on the existing generalization degree of the applied predictive models.