| Title |
CNN based Reinforcement Learning for Driving Behavior of Simulated Self-Driving Car |
| Authors |
조영완(Youngwan Cho) ; 이종석(Jongseok Lee) ; 이광엽(Kwangyup Lee) |
| DOI |
https://doi.org/10.5370/KIEE.2020.69.11.1740 |
| Keywords |
Self-Driving; Driving behavior learning; Reinforcement learning; DDPG; CNN |
| Abstract |
This paper proposes a self-learning method for autonomous vehicle driving behavior using reinforcement learning without considering the dynamic model of the vehicle. In order to make decision needed for determine the optimal driving behavior (steering, throttle, brake) to achieve a given driving purpose in each state by using state information of the vehicle, such as vehicle movement speed, direction, degree of deviation from the center of the track, and distance to the edge of the track, we propose a method of applying the reinforcement learning by the DDPG structure and further using the driving image to improve driving performance. In this paper, we propose structures of an action decision network(Actor) and an action value evaluation network(Critic) to implement the DDPG learning model. We also propose a prediction model for predict the next state driving image based on the current driving image to improve driving performance in the corner path and a corner classifier for classifying the driving track type. The method proposed in this paper was implemented in a TORCS simulator environment, and the performance of the target driving behavior was evaluated through applying the learning model to driving agent. |