A Traffic Prediction Enabled Double Rewarded Value Iteration Network for Route Planning

Effective route planning is the key to improving transportation efficiency. By leveraging the in-depth knowledge of road topology and traffic trends, experienced drivers (e.g., taxi drivers) can usually find near-optimal routes. However, existing online route planning services can hardly acquire this domain knowledge, so they just provide the fastest/shortest route based on current traffic conditions. These seemingly optimal routes may attract numerous vehicles and then become extremely congested. To solve this problem and actually improve transportation efficiency, we propose a double rewarded value iteration network (VIN) to fully learn the experienced drivers' routing decisions, which are based on their implicitly estimated traffic trends. First, the global traffic status and routing actions are chronologically extracted from large-scale taxicab trajectories. Then, to model the knowledge of traffic trends, a long short-term memory network is trained. Being expert at learning long-term planning involved functions, the VIN is utilized to model the policy function from both current and predicted future traffic status to an experienced driver's routing action. Finally, the performance of our proposed model is evaluated on real map and taxicab trajectories in Beijing, China. The experimental results demonstrate that the proposed model can achieve human like performance in most cases, with high success rate and less commuting time.