Deep Reinforcement Learning for Autonomous Driving

« Artificial Intelligence is a vast topic often confusing for an uninitiated person. In this paper, we describe a technique of Machine Learning which is Reinforcement Learning. Surely the less known but also the most spontaneous kind of Artificial Intelligence. The idea is to learn through experience by defining a reward for each action. Mathematically, it is based on a Markov Decision Process leading an agent (autonomous car for example) to make the best action. A problem of Reinforcement Learning is that it needs information on the environment and it is often too large.

The solution is to use Deep Learning to extract essential features of the environment. That learning is based on Artificial Neural Networks.

Thanks to a literature review of artificial neural networks, deep learning and reinforcement learning, we are able to address the challenges related to autonomous driving, especially that of recognizing the environment to protect the occupants of the car and other users. »

Index terms

Artificial Intelligence: the process of imitating human intelligence based on the creation and application of computer algorithms

Machine learning: Domain of AI where computer algorithms learn by themselves and correct their mistakes

Deep Learning: Machine learning technique based on Artificial Neural Networks

Reinforcement Learning: Domain of machine learning that uses the environment to learn how to react to different situations

Artificial Neurons: Inspired by an animal neuron, it receives inputs that it processes through a mathematical function. Then it sends back or not an output

Markov Decisions process: Stochastic model where an agent makes moves between states only according to the current one

Autonomous driving: Driving is said to be autonomous when a system can assist or even replace the driver. This includes simple speed control up to full control of the vehicle

Q-function: Function that takes as parameters a state and an action and returns the value of that action given that state. This value allows to compare actions under a policy

Policy: Strategy that allows selecting the best action given a state

LIDAR (light detection and ranging): A remote sensing and telemetry method that emits pulses of infrared light and then measures their return time after being reflected off nearby objects