Enabling Self-defined Navigation on Road Graph via Double Rewarded Generalized VIN

摘要

Nowadays navigation systems are mostly centralized and mainly relying on the traffic conditions of road network, and other influence factors are rarely considered, such as roads topology, domain knowledge and planning autonomy, which is result in the recommended seemingly optimal routes may attract numerous vehicles and then congestion. Besides, the existing region-based methods have deviation on acquiring the traffic conditions of road segments, which lead to the inadequate use of road resource and the waste of drivers’ time.Considering that MEC can provide low latency and quick responses support for intelligent driving, we propose an online planning approach on the road graph via a double rewarded Generalized Value Iteration Network, 2r-GVIN for short. In detail, a large-scale real GPS trajectories are mapped and discretized into traffic sequence pairs anonymously, that is, capturing the traffic status and driving actions as dataset. Then a neural network ConvLSTM which can effectively capture the spatio-temporal properties, is trained to predict the future traffic situation to form “prediction reward”. To enable learning and planning on irregular road graphs straightforwardly, an economical embedding graph convolution operator network GVIN is used. Combining with the model input “current reward”, 2r-GVIN is formed to solve the plan-involved function iteratively. Lastly, we train and evaluate 2r-GVIN on planning problem in road networks, showing that it can achieve road segment-based self-defined navigation with high success rate.