AWS DeepRacer

Abstract

AWS DeepRacer League is the world's global autonomous racing league. In the league, partitions will train their own AI models to drive the cars and win the competition. In this project, reinforcement learning is used, and Proximal Policy Optimization (PPO) Algorithm is applied for model training. Since the action spaces and the reward function are the essential elements which affect the model performance. Therefore, the project will focus on how to implement them to achieve a better result.

To improve the action spaces, the optimal racing line for the Penbay circuit, which is the racing track of the March qualifier, was calculated with the K1999 path optimization algorithm's inspiration. After that, by collecting the steering angles and the velocity through the track, Kmean clustering is applied to select the action spaces.

Furthermore, hyperparameters are another factor that will significantly affect the training result. Therefore, during the model training, log analysis will be performed for evaluating those three elements. For the final result, the model has ranked 6 in the March qualifier.