Adaptive Path Planning for Autonomous Vehicles in Complex Traffic Scenarios

Over the past decades, an Autonomous Vehicles (AVs) driving functions have obtained the more interests from both the industry and academia. The AVs needed a basic planner for all the practicable scenarios and the recent research develops named a planner through the integrated scenario interpretation. Nevertheless, it may remarkably enhance the planner difficulty indeed in some simple operations. Examining the safety of AVs is the significant challenge in relation to their distribution on the roads because of large number of achievable conditions happened in traffic. In this research, the adaptive path planning approach of AV established multi-light trained Reinforcement Learning (RL) is proposed, aimed to enhance the fuel cost as well as AV comfort. By utilizing a logical Deep Q-learning function, the training approach obtains the key setting data collected through the vehicles as inputs, eventually outcomes an acceleration that enhances a cumulative reward. The proposed method attains better results by using the performance metrices of average acceleration, distance and computational time values about 0.149, 0.98 and 0.49 respectively compared with Pertinent Boundary-based Unified Decision (PBUD), Adaptive Mode Predictive Control (MPC) and Dynamic Decoupling Lane-Changing (DDLC).