Multi-agent based vehicle routing

Abstract

Traffic congestion is a menace in the society with serious economic implications. Over the years, several approaches have been developed to mitigate the issue and plan better routes for the vehicles. This problem has been tackled both on a System level and driver level perspectives. This Final Year Project Report is a summarization of the work done by the author over his FYP-URECA projects. This report has been split into two parts with Part I summarizing the node pressure minimization problem solved using a novel MAS and Part II is aimed at explaining the improvement in design and implementation of the Stochastic Routing System developed. Pressure based routing systems are known to be effective in wireless sensor networks, however rarely applied to transportation systems due to the specific domain knowledge. Part I of this report summarizes the methodology used in adapting the node pressure concept into the traffic management. In particular, it discusses a Multi-Agent System (MAS) with vehicle agents and infrastructure agents, which can respectively provide static and dynamic solutions to reduce the node pressure.Pertaining to the static solution, the infrastructure agents rely on a reinforcement learning method to calculate the optimal routes for vehicle agents. In the dynamic solution, the infrastructure agents dynamically adjust and re-route vehicle agents based on a novel multi-unit combinatorial auctioning system proposed. Extensive experiments on realistic traffic simulation platform have proven our methods, especially the dynamic solution, to achieve significant improvement in the reduction of node pressure and travel-times for the vehicle agents in comparison to other approaches. Part II of this report is aimed at summarizing the design improvements to the existing stochastic routing system, which is in the state of continuous development. The proposed design improvements were implemented into the system and was tested using the BMW i3 car.