Machine Learning-Based Minimization of Empty Truck Repositioning

Abstract

This study applies a Deep Reinforcement Learning (DRL) algorithm for a multi-agent system integrated within a Markov decision process framework to tackle the empty repositioning problem of the cargo truck fleet. The goal is to learn a relocation strategy that makes supply-demand aware actions under stochastic customer demands to reduce repositioning costs and the associated CO2 emissions and to provide a decisionsupport tool for short-term operational planning. Further, the performance of the obtained policy is evaluated and compared with a designed rule-based policy with tunable parameters against significant key performance indicators. This is accomplished by implementing a simulated environment utilizing a real Geographic Information System (GIS) map of the UAE to replicate the dynamics of the problem. According to the experimental results, the reinforcement learning policy achieved a 5% reduction in cost and a 9% decrease in the waiting time of orders for a small problem setting compared to the rule-based policy performance. It also improved the customer demand fulfillment rate, demonstrating a potential for applying the developed DRL model to address larger-scale problems.

Date of Award	Jul 2022
Original language	American English