Autonomous Car: Deployment of Reinforcement Learning in Various Autonomous Driving Applications

EasyChair Preprint 1305

Abstract

Reinforcement Learning is a type of Machine Learning which allows machines and software agents to automatically determine the ideal behavior in order to maximize their performance. The possible applications of Reinforcement Learning are many and in particular ranges from controlling vehicle to find the most efficient motor combination, to autonomous car navigation where collision avoidance behavior can be learnt by negative feedback from bumping into obstacles. For a vehicle to operate autonomously several real-time systems must work together and these include environment mapping and understanding, localization, route planning and movement control. An overview of different reinforcement learning applications in Autonomous Driving systems is presented. The deep reinforcement learning in single agent setting using convolutional neural networks with Q-Learning and how the single-agent model can be used to produce the specific driving behaviour of an autonomous car on a highway is applied. For this, autonomous cars are considered as agents learning to drive safely and a traffic simulator is created – including fixed agents and human drivers – that serves as the learning environment. Finally, results show that the model using neural networks in a single-agent setting perform well when the traffic density is lower.

Keyphrases: Autonomous Car, Deep Reinforcement Learning, Reinforcement Learning, driverless car

@booklet{EasyChair:1305,
  author    = {Poondru Prithvinath Reddy},
  title     = {Autonomous Car: Deployment of Reinforcement Learning in Various Autonomous Driving Applications},
  howpublished = {EasyChair Preprint 1305},
  year      = {EasyChair, 2019}}