Synopsis Time-dependent Routing by : Rabie Jaballah
The vehicle routing problem (VRP), introduced more than 60 years ago, is at the core of transportation systems. With decades of development, the VRP is one of the most studied problems in the literature, with a very rich set of variants. Yet, primarily due to the lack of data, two critical assumptions make the VRP fail to adapt effectively to traffic and congestion. The first assumption considers that the travel speed is constant over time ; the second, that each pair of customers is connected by an arc, ignoring the underlying street network. Traffic congestion is one of the biggest challenges in transportation systems. As traffic directly affects transportation activities, the whole supply chain needs to adjust to this factor. The continuous growth of freight in recent years worsens the situation, and a renewed focus on mobility, environment, and city logistics has shed light on these issues. Recently, advances in communications and real-time data acquisition technologies have made it possible to collect vehicle data such as their location, acceleration, driving speed, deceleration, etc. With the availability of this data, one can question the way we define, model, and solve transportation problems. This allows us to overcome the two issues indicated before and integrate congestion information and the whole underlying street network. We start by considering the whole underlying street network, which means we have customer nodes and intermediate nodes that constitute the street network. Then, we model the travel time of each street during the day. By dividing the day into small intervals, up to a precision of a second, we consider precise traffic information. This results in a new problem called the time-dependent shortest path vehicle routing problem (TD-SPVRP), in which we combine the time-dependent shortest path problem (TD-SPP) and the time-dependent VRP (TD-VRP), creating a more general and very challenging problem. The TD-SPVRP is closer to what can be found in real-world conditions, and it constitutes the topic of Chapter 2, where we formulate it as a mixed-integer linear programming model and design a fast and efficient heuristic algorithm to solve this problem. We test it on instances generated from actual traffic data from the road network in Québec City, Canada. Results show that the heuristic provides high-quality solutions with an average gap of only 5.66%, while the mathematical model fails to find a solution for any real instance. To solve the challenging problem, we emphasize the importance of a high-performance implementation to improve the speed and the execution time of the algorithms. Still, the problem is huge especially when we work on a large area of the underlying street network alongside very precise traffic data. To this end, we use different techniques to optimize the computational effort to solve the problem while assessing the impact on the precision to avoid the loss of valuable information. Two types of data aggregation are developed, covering two different levels of information. First, we manipulated the structure of the network by reducing its size, and second by controlling the time aggregation level to generate the traffic data, thus the data used to determine the speed of a vehicle at any time. For the network structure, we used different reduction techniques of the road graph to reduce its size. We studied the value and the trade-off of spatial information. Solutions generated using the reduced graph are analyzed in Chapter 3 to evaluate the quality and the loss of information from the reduction. We show that the transformation of the TD-SPVRP into an equivalent TD-VRP results in a large graph that requires significant preprocessing time, which impacts the solution quality. Our development shows that solving the TD-SPVRP is about 40 times faster than solving the related TD-VRP. Keeping a high level of precision and successfully reducing the size of the graph is possible. In particular, we develop two reduction procedures, node reduction and parallel arc reduction. Both techniques reduce the size of the graph, with different results. While the node reduction leads to improved reduction in the gap of 1.11%, the parallel arc reduction gives a gap of 2.57% indicating a distortion in the reduced graph. We analyzed the compromises regarding the traffic information, between a massive amount of very precise data or a smaller volume of aggregated data with some potential information loss. This is done while analyzing the precision of the aggregated data under different travel time models, and these developments appear in Chapter 4. Our analysis indicates that a full coverage of the street network at any time of the day is required to achieve a high level of coverage. Using high aggregation will result in a smaller problem with better data coverage but at the cost of a loss of information. We analyzed two travel time estimation models, the link travel model (LTM) and the flow speed model (FSM). They both shared the same performance when working with large intervals of time (120, 300, and 600 seconds), thus a higher level of aggregation, with an absolute average gap of 5.5% to the observed route travel time. With short periods (1, 10, 30, and 60 seconds), FSM performs better than LTM. For 1 second interval, FSM gives an average absolute gap of 6.70%, while LTM provides a gap of 11.17%. This thesis is structured as follows. After a general introduction in which we present the conceptual framework of the thesis and its organization, Chapter 1 presents the literature review for the two main problems of our development, the shortest path problem (SPP) and the VRP, and their time-dependent variants developed over the years. Chapter 2 introduces a new VRP variant, the TD-SPVRP. Chapter 3 presents the different techniques developed to reduce the size of the network by manipulating spatial information of the road network. The impact of these reductions is evaluated and analyzed on real data instances using multiple heuristics. Chapter 4 covers the impact of time aggregation data and travel time models when computing travel times on the precision of their estimations against observed travel times. The conclusion follows in the last chapter and presents some research perspectives for our works.