Deep imitation learning: the impact of depth on policy performance

This paper investigates the impact of network depth on the performance of imitation learning applied in the development of an end- to-end policy for controlling autonomous cars. The policy generates optimal steering commands from raw images taken from cameras attached to the car in a simulated environment. A convolutional neural network (CNN) is used to find the mapping between inputs (car images) and the desired steering angle. The CNN architecture is modified by changing the number of convolutional layers as well as the filter size. It is observed that the learned policy is capable of driving the car in the autonomous mode purely using visual information. In addition, simulation results indicate that deeper CNNs outperform shallower CNNs for learning and mimicking the human driver’s behavior. Surprisingly, the best performance is not achieved by the most complex CNN.