Real-time deep reinforcement learning based vehicle navigation

Traffic congestion has become one of the most serious contemporary city issues as it leads to unnecessary high energy consumption, air pollution and extra traveling time. During the past decade, many optimization algorithms have been designed to achieve the optimal usage of existing roadway capacity in cities to leverage the problem. However, it is still a challenging task for the vehicles to interact with the complex city environment in a real time manner. In this paper, we propose a deep reinforcement learning (DRL) method to build a real-time intelligent vehicle routing and navigation system by formulating the task as a sequence of decisions. In addition, an integrated framework is provided to facilitate the intelligent vehicle navigation research by embedding smart agents into the SUMO simulator. Nine realistic traffic scenarios are simulated to test the proposed navigation method. The experimental results have demonstrated the efficient convergence of the vehicle navigation agents and their effectiveness to make optimal decisions under the volatile traffic conditions. The results also show that the proposed method provides a better navigation solution comparing to the benchmark routing optimization algorithms. The performance has been further validated by using the Wilcoxon test. It is found that the achieved improvement of our proposed method becomes more significant under the maps with more edges (roads) and more complicated traffics comparing to the state-of-the-art navigation methods.