Choosing an appropriate route is a critical problem in urban navigation. Being familiar with roads topology and other vehicles’ routes, experienced drivers could usually find a near optimal route. However, the nowadays navigation applications hardly catch the domain knowledge and drivers’ interaction in this scenario. In fact, they only recommend several routes and leave the most difficult decision to driver. Hence the route is often congested by many vehicles whose drivers make a similar choose. To intelligently make right decision on navigation and improve traffic efficiency for each vehicle, we propose a neural network structure which learns to plan coarse-grained route in complex urban areas.Focused on learning to navigate, this paper first formalizes urban map and vehicle route model. The city map is segmented into grids and each vehicle’s route is mapped to grids. Based on grid-world model, we solicit both global traffic status and driving actions from large-scale taxicab GPS data. The learn-to-plan problem is therefore to find a policy function from a global status representation to an experienced driver’s action under that status. Traditional neural network is difficult to learn to this plan-involved function, so we leverage and modify value iteration network (VIN), which explicitly takes long-term plan into consideration. Finally we evaluate the performance of proposed network on real map and trajectory data in Beijing, China. The results show that VIN can achieve human driver performance in most cases, with high success rate and less commuting time.


    Zugriff

    Zugriff prüfen

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    Learning Urban Navigation via Value Iteration Network


    Beteiligte:
    Yang, Shu (Autor:in) / Li, Jinglin (Autor:in) / Wang, Jie (Autor:in) / Liu, Zhihan (Autor:in) / Yang, Fangchun (Autor:in)


    Erscheinungsdatum :

    2018-06-01


    Format / Umfang :

    3333742 byte




    Medientyp :

    Aufsatz (Konferenz)


    Format :

    Elektronische Ressource


    Sprache :

    Englisch



    Comparison of Value Iteration, Policy Iteration and Q-Learning for solving Decision-Making problems

    Hamadouche, Mohand / Dezan, Catherine / Espes, David et al. | IEEE | 2021


    AUTONOMOUS SOARING POLICY INITIALIZATION THROUGH VALUE ITERATION

    Rothaupt, Benjamin J. / Notter, Stefan / Fichter, Walter | TIBKAT | 2021


    Autonomous Soaring Policy Initialization Through Value Iteration

    Rothaupt, Benjamin J. / Notter, Stefan / Fichter, Walter | AIAA | 2021



    INavFIter-M: Matrix Formulation of Functional Iteration for Inertial Navigation Computation

    Jiang, Hongyan / Zhu, Maoran / Fu, Yanyan et al. | TIBKAT | 2023