Improved Q-Learning Algorithm for Path Planning of Autonomous Driving

Slow convergence and low efficiency in learning limit the applications of the path planning algorithms. In order to address these issues, this paper presents an improved Q-learning algorithm. Firstly, it dynamically adjusts the exploration factor which enables the agent to prioritize exploration during the early stages of path planning and to use the explored environmental information more effectively in the middle and later stages, thereby improving the convergence speed of the algorithm. Subsequently, the process of repeatedly learning about the same obstacle is reduced by employing an obstacle learning matrix and leading to an increased success rate of the algorithm. Lastly, the algorithm’s performance is further improved by increasing the action step size. Simulation experiments have shown that the proposed improved Q-learning algorithm not only improves convergence speed and learning efficiency but also shortens the length of path planning by comparing with the traditional ones.