Optimization methods incorporating statistical and machine learning models in traffic safety

Abstract

In this dissertation, we investigated the optimization and statistical models regarding the safety aspects of trucking operations. A significant gap in the literature has been identified concerning the interaction between these two kinds of models. The majority of the optimization models in the discipline of transportation safety are related to hazardous transportation with the specialty of high consequence and low probability. For general risk on the road and truck drivers, statistically significant risk factors have been extensively used in optimization approaches. The first chapter is concerned with describing the relevant literature and corresponding gap. To address this disparity, a dynamic model that is easy to combine with the statistical risk modeling approaches has been developed to help the driver schedule the rest stops and select the optimal speed during the route under different conditions in the second chapter. Here, the objective of the dynamic model is to minimize the cumulative travel time and the risk on the road simultaneously by using the averaged weighted method to combine the two objectives. Another approach to incorporating risk factors into trucking routing, referred to as the bi-objective k-shortest paths problem has been studied in the third chapter. This method's framework comprises three major steps: finding the k-shortest paths, predicting the total travel time and risk for each path, and applying Pareto ranking to find the non-dominate sets of paths. The risk and travel time are time-related factors due to time-dependent weather and speed. With the help of the bi-objective model, the driver can determine the optimal route by taking into account multiple factors. Both approaches in dynamic model and bi-objective ksp model rely on explicitly constructed risk models, arrived at through either statistical or machine learning modeling. Another possible framework for incorporating driving-related data into decision-making is inverse reinforcement learning (IRL). IRL is a method to allow us to learn from the expert's data from decision-making. More specifically, by building a Markov Decision Processes (MDP) model, the purpose of IRL is to find the reward for each state-action pair. The reward is structured according to those features variables related to the state and action variables in the MDP model. Thus, it gives us an opportunity to understand how those experts determine the best action given various driving conditions. The results show that the weather features are significant factors in determining the mean speed. Consequently, the derived reward function can be employed in a Q-learning framework to determine the optimal policy.