Vehicle Routing Problem with Value Iteration Network

Abstract

Multi Vehicle routing to service consumers in dynamic and unpredictable surrounding such as congested urban areas is a difficult operation that needs robust and flexible planning. Value iteration networks hold promise for planning in vehicle routing problem. Conventional approaches aren’t usually constructed for real-life settings, and they are too slow to be useful in real-time. In comparison, Vehicle Routing Problem with Value Iteration Network (VRP-VIN) offers a neural network model based on graphs that can execute multi-agent routing in a highly dispersed but connected graph with constantly fluctuating traffic conditions using learned value iteration. Furthermore, the model’s communication module allows vehicles to work better in a cooperative manner online and can easily adapt to changes. A virtual environment is constructed to simulate real-world mapping by self-driving vehicles with uncertain traffic circumstances and minimal edge coverage. This method beats standard solutions based on overall cost and run-time. Experiments show that the model achieves a total cost difference of 3% when compared with a state-of-art solver having global information. Also, after being trained with only 2 agents on networks with 25 nodes, can easily generalize to a scenario having additional agents (or nodes).