Explainable navigation system using fuzzy reinforcement learning

  • PDF / 1,970,647 Bytes
  • 18 Pages / 595.276 x 790.866 pts Page_size
  • 61 Downloads / 269 Views

DOWNLOAD

REPORT


ORIGINAL PAPER

Explainable navigation system using fuzzy reinforcement learning Rolando Bautista-Montesano1

· Rogelio Bustamante-Bello1 · Ricardo A. Ramirez-Mendoza1

Received: 17 July 2020 / Accepted: 22 September 2020 © Springer-Verlag France SAS, part of Springer Nature 2020

Abstract Abstract Explainable outcomes in autonomous navigation have become crucial for drivers, other vehicles, as well as for pedestrians. Creating trustworthy strategies is mandatory for the integration of self-driving cars into quotidian environments. This paper presents the successful implementation of an explainable Fuzzy Deep Reinforcement Learning approach for autonomous vehicles based on the AWS DeepRacerTM platform. A model of the environment is created by transforming crisp values into linguistic variables. A fuzzy inference system is used to define the reward of the vehicle depending on its current state. Guidelines to define the actions and to improve performance of the reinforcement learning agent are given based on the characteristics of the existing hardware. The performance of the models is tested on tracks with distinctive properties using agents with different policies and action spaces, and shows explainable and successful navigation of the agent on diverse scenarios. Graphic Abstract

Keywords Deep learning · Fuzzy inference system · Explainable artificial intelligence · Reinforcement learning · Autonomous vehicle

Abbreviations AWS Amazon Web Services AI Artificial intelligence XAI Explainable artificial intelligence

B B

Rolando Bautista-Montesano [email protected]

Fuzzy inference system Membership function Machine learning Deep learning Deep neural network Convolutional neural network Reinforcement learning

Ricardo A. Ramirez-Mendoza [email protected] Rogelio Bustamante-Bello [email protected]

1

FIS MF ML DL DNN DNN RL

Tecnologico de Monterrey, Campus Ciudad de Mexico, Calle del Puente 222, Col. Ejidos de Huipulco, Tlalpan, CDMX, Mexico

1 Introduction The automotive academic and industrial research and development changed drastically in 2005 when the Defense Advanced Research Projects Agency (DARPA) Grand Chal-

123

International Journal on Interactive Design and Manufacturing (IJIDeM)

lenge was won by Stanford’s Stanley [1]. This event opened an enormous range of opportunities in diverse fields such as computer science, electrical, mechanical, transportation and logistics engineering branches. Due to the interdisciplinary nature of the problem, the transfer of technology and knowledge between academic communities and industry has become vital to the development of a solution. Academically developed prototypes of self-driving cars systems were required to fit industry standards in terms of performance, safety, and trust [2]. The autonomous navigation problem seemed to be solved after the DARPA Urban Challenge [3], however, new challenges arose. These challenges comprise the trending research topics: data management, adaptive mobility, cooperative planning, validation, and evaluation. Data uncerta