Evaluating Architecture Impacts on Deep Imitation Learning Performance for Autonomous Driving

Imitation learning has gained huge popularity due to its promises in different fields such as robotics and autonomous systems. A great deal of past research work in the field of imitation learning has been devoted to developing efficient and effective policies using deep convolutional neural networks (CNNs). The performance of CNN-based control policies intimately depends on the network architecture. Determination of the optimal architecture for CNNs is still a hot research topic for the deep learning community. This study comprehensively investigates and quantifies the impact of CNN architecture on the performance of learned policy for an autonomous vehicle. CNN models with different architectures (number of layers and filters) are fed by visual information from multiple cameras obtained from multiple driving simulations. These networks are trained to precisely find the mapping between visual information and the steering angle. Two ensemble approaches are also introduced to further improve the overall accuracy of steering angle estimations. Obtained results indicate that deeper networks show a better performance than less deep networks during autonomous driving. Also it is observed that best results are achieved by ensemble approaches.