A comparison study of deep neural controllers and classic controllers in self-driving car application

In this paper deep neural controller is evaluated in self-driving car application which is one of the most important and critical among human-in-the-loop cyber physical systems. To this aim, the modern controller is compared with two classic controllers, i.e. proportional–integral–derivative and model predictive control for both quantitative and qualitative parameters. The parameters reflect three main challenges: (i) design-time challenges like dependency to the model and design parameters, (ii) implementation challenges including ease of implementation and computation workload, and (iii) run-time challenges and parameters covering performance in terms of speed, accuracy, control cost and effort, kinematic energy and vehicle depreciation. The main objective of our work is to present comparison and concrete metrics for designers to compare modern and traditional controllers. A framework for design, implementation and evaluation is presented. An end-to-end controller, constituting six convolution layers and four fully connected layers, is evaluated as the modern controller. The controller learns human driving behaviors and is used to drive the vehicle autonomously. Our results show that despite the main advantages of the controller i.e. being model free and also trainable, in terms of important metrics, this controller exhibits acceptable performance in comparison with proportional–integral–derivative and model predictive controllers.