Highway Autopilot Using Deep Reinforcement Learning and Graph Neural Networks

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Date

2023-12-11

Department

Major/Subject

Autonomous Systems

Mcode

ELEC3055

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

46+1

Series

Abstract

Autonomous vehicles hold promise for enhancing road traffic safety by minimizing human errors through optimal decision-making approaches. The development of decision-making algorithms for complex traffic scenarios is crucial for the successful implementation and wider adoption of autonomous vehicles. In the context of a multi-laned highway traffic scenario, lane changing is a challenging coordination problem due to high speeds and the variation in the number of surrounding road users. Further, a robust input representation of the highway traffic is required for the autonomous vehicle to make optimal decisions for safe navigation. Learning-based or data-driven methods offer a promising solution for developing these robust decision-making algorithms. In this thesis, the feasibility of developing a low-level controller model based on deep reinforcement learning and graph neural networks which can be used to safely change lanes in a highway with traffic in a simulated environment is investigated. The dependence on the input graph traffic configuration on the model’s performance is shown for a given simulation environment configuration. Further, the usefulness of the graph attention mechanism for understanding the decision-making of the trained policy is explored.

Description

Supervisor

Kyrki, Ville

Thesis advisor

Laine, Leo

Keywords

deep reinforcement learning, graph neural networks, soft actor-critic, graph attention network, multi-laned Highway

Other note

Citation