Multi-Agent Reinforcement Learning For Multi Vehicles One-commodity Vehicle Routing Problem
Yamen Habib, Andrey Filchenkov
Abstract
Vehicle Routing Problem with Dynamic and Stochastic information (DS-VPR) is mathematical problem reflecting a huge set of real-world logistic problem. In this paper, we presented a novel method to handle DS-VRP. Our approach depends on multi-agent reinforcement learning where we place our decision makers in nodes instead of vehicles and we use geometric Laplacian eigenmaps embedding to represent graph nodes information. We also presented a new method of training RL agents in a graph by separating it into two phases. Also, we built a simulation to develop and test different methods for the DS-VRP problem. The results showed that our approach shows the same performance being trained from the scratch as Google OR tools and MARDAM, while being more adaptive than these methods and having less parameters than MARDAM.