Adaptive Cruise Control Using Twin Delayed Deep Deterministic Policy Gradient

Abstract

With the emergence of different machine learning and high-performance systems, the field of reinforcement learning has gained a plethora of interest from industry as well as researchers in the academic field. In this work, an algorithm known as twin delayed deep deterministic policy gradient (TD3) that uses critic and actor networks is implemented to provide an efficient solution to the adaptive cruise control (ACC) problem. A reward function is used for the same that considers error in velocity and control input as parameters in addition to an extra term to give necessary push to reward function in case the error is less than a particular value. Two test cases are considered to substantiate the working effectiveness of the algorithm proposed. In one case normal driving scenario is considered while in the other test case a disturbance is modelled to replicate actual condition. From the results of the simulation and analysis, the operational efficacy of the put forward twin delayed deep deterministic policy gradient algorithm is found out.