Deep Reinforcement Learning for Shared Autonomous Vehicles (SAV) Fleet Management

TL;DR

This paper investigates reinforcement learning strategies to optimize shared autonomous vehicle fleet management, aiming to reduce rider wait times, parking costs, and empty cruising, addressing key challenges in SAV deployment.

Contribution

It introduces novel reinforcement learning approaches tailored for SAV fleet management, focusing on minimizing multiple operational costs simultaneously.

Findings

RL approaches effectively reduce rider waiting times

Proposed methods decrease parking space costs

Strategies cut down empty cruising time

Abstract

Shared Automated Vehicles (SAVs) Fleets companies are starting pilot projects nationwide. In 2020 in Fairfax Virginia it was announced the first Shared Autonomous Vehicle Fleet pilot project in Virginia. SAVs promise to improve quality of life. However, SAVs will also induce some negative externalities by generating excessive vehicle miles traveled (VMT), which leads to more congestions, energy consumption, and emissions. The excessive VMT are primarily generated via empty relocation process. Reinforcement Learning based algorithms are being researched as a possible solution to solve some of these problems: most notably minimizing waiting time for riders. But no research using Reinforcement Learning has been made about reducing parking space cost nor reducing empty cruising time. This study explores different \textbf{Reinforcement Learning approaches and then decide the best approach to help minimize the rider waiting time, parking cost, and empty travel