DeepRacer Model Training for autonomous vehicles on AWS EC2

Autonomous vehicle (AV) is the future of public transportation to reduce the road congestion and accidents. However, fully self-driving car is still a challenge for carmaker, and they must ensure the maximum driving security to avoid the ethical issue. To develop a mature model to AV, reinforcement learning becomes a solution which can explore many possibilities and choose the best possible action choice facing different road conditions. AWS DeepRacer is a comprehensive platform for researchers to start reinforcement learning for AV. With using DeepRacer Console, all the parameters including action spaces, reward functions and hyper parameters can be edited online and trained remotely. However, the price for training a converged model on DeepRacer Console is quite expensive for beginners. This paper present a DeepRacer simulation process built on EC2 instance which demand no hardware and cost significantly less than regular DeepRacer Console. Based on the default models from AWS, the authors experimentally adjusted hyper parameters and added reward functions to achieve higher speed and smoother driving actions. Even though the computing resources still limited the agent performance and stopped the model from convergence, there are 2-3m/s speed increases when adding low speed penalty and progress penalty on reward function.