Routing algorithms for optimisation of transportation systems

Abstract

As efficiency and sustainability become more important, researchers are working on new concepts to improve the way logistics are handled in the current Supply Chain Management (SCM). One concept gaining popularity is the delivery of products in urban areas using bicycles. More companies started using bicycles as an alternative transportation mode and face challenges to efficiently satisfy their customers and employees needs. Large cities with uphill roads require transportation systems to take into account the energy needed by cyclists to move. The load carried on push-bikes has to be kept under a certain threshold for cyclists to be able to pedal on ascending roads. Therefore, cyclists have to choose their route differently to vehicle drivers, hence the need for optimised routing. Another concept gaining popularity is the collaboration in logistics between several companies. Collaboration is thought to be an enabler for a sustainable and efficient SCM. One important improvement of modern logistics is the frequent exchange of containers via multiple cross-docks which requires spatial and time synchronisation between different types of vehicles. Since each logistics network has its specificities and requirements, new solutions such as the Physical Internet (PI) arise with standardised PI-Containers and protocols to face this challenge. By connecting several transportation networks through collaboration, the PI is expected to considerably improve the way logistics are handled in the current SCM. Connecting several distribution networks will produce continually varying network conditions arising from traffic growth. Some regions of the network would suffer from bottlenecks that could lead to an increase in transportation costs. One reason is that traditional routing protocols in logistics do not learn from their previous experiences of network problems such as congestion. Therefore, an intelligent network traffic control method is essential to avoid this problem. As routing problems are at centre stage in transportation sciences, they represent a critical research scope to study in order to improve logistics. As a consequence, in this thesis, three research directions involving routing problems are identified and solved to provide more flexible and extended models to the aforementioned research gaps. First, a new problem is introduced to tackle constraints arising for bicycle deliveries. A novel Mixed-Integer Linear Programming (MILP) model and an Evolutionary Local Search algorithm are developed to efficiently solve the problem. Experimental results show the accuracy and stability of the proposed algorithm compared to the CPLEX solver from IBM on a wide range of generated instances. A real-world scenario is also studied to demonstrate the relevance of this method. Second, this research focuses on the way the VRP can be solved while considering an important number of attributes for real-life applications. A rich vehicle routing problem with pickup and delivery including several attributes for the PI is studied. A mathematical formulation is proposed and implemented in CPLEX to solve the problem. The model is then extended to handle multi-objective, uncertainty and dynamism. Multi-threaded meta-heuristics based on Simulated Annealing and Genetic Algorithm are developed with a set of new operators to handle the problem specificities. Computational results on a generated data-set showed that the proposed meta-heuristics are superior to CPLEX in terms of solvability and computational time. A classical benchmark on pickup and delivery problems was also used to validate the proposed method against state-of-the-art methods. The algorithms are integrated into a framework and used to provide solutions for a local company. Third, the paradigm between the digital and physical internet is explored to propose a new routing approach based on packet routing. While deep learning has been demonstrated to be a promising method for solving numerous optimisation problems efficiently, its application on packet routing is relatively new and rare. This research provides a proof of concept and sheds light on new opportunities to design efficient routing protocols for logistics with deep learning. Simulation results demonstrate that the proposed method is able to not only learn the shortest path for deliveries but also to take into account the truck fulfilment rates to improve global efficiency.