Autonomous vehicle driving via deep deterministic policy gradient

Autonomous driving has became one of the most hot trends in artificial intelligence area in recent years thanks to the machine learning algorithms. However, most of the autonomous driving studies are still limited to discrete action space. In this study, we propose to implement Deep Deterministic Policy Gradient algorithm for learning driving behavior over the continuous actions. For this purpose, a driving simulator is employed which interfaces with IPG CarMker software where the virtual environment and dynamical vehicle model can be built.”Human-in-the-loop” is performed in order to gather the data and a neural network which is implemented in Behavior Layer is trained to recognize two different scenarios-forward driving and stop. Based on the scenario the agent is dealing with, the actions are learnt and suggested from the DDPG algorithm. The experimental results show that DDPG algorithm is able to learn the optimal policy with continuous actions reliably for both scenarios.