Autonomous Driving Systems: Developing an Approach based on A* and Double Q-Learning

Autonomous driving is the most attractive field to research by academic and industrial socials that intelligent transportation play a vital role in structure of autonomous driving systems. Artificial Intelligence (AI) is an infrastructure for autonomous driving by designing of intelligent machine. Deep Learning is one of subfields of Artificial Intelligence that create models by mimicking human brain’s functioning to make decision that it has shown great success in autonomous diving systems field. However, it performs very poorly in some stochastic environments caused by large overestimations of action values. Thus, we use the double estimator to Q-learning to construct Double Q-learning with a new off-policy reinforcement learning algorithm. By this algorithm, we can approximate the maximum expected value for any number of random variables and it underestimate rather than overestimate the maximum expected value. Moreover, we use an optimization method based on A* to improve routing in automation driving. Our proposed approach based on double Q-Learning and A* is evaluated on an example environment with random obstacles and compare results to use Q-Learning alone. Results show the proposed approach has better performance based on duration of trip to destination and collision to obstacles.