Creation of real world test environments and scenarios for traffic flow research is impossible and therefore sophisticated simulators should be used for the task. CARLA (Dosovitskiy et al. 2017) is an open source, flexible urban driving simulator for aiding the research on automated vehicles. However, for our purposes the simulator is too simplistic at its present state. Thus, the research will include:
In order to create a tool providing advise for planning the traffic in city planning procedure, the simulator should contain realistic representation of the planned city area layouts. Thus, a city model is created into CARLA based on area plans of selected scenarios.
At present, CARLA enables simulation of the traffic flow at a microscopic level, namely simulates the movement of individual vehicles and generates other actors in the traffic using standard Unreal Engine’s vehicle model and their motion using a basic controller defining their behavior. To achieve our research goals we need to have more intelligence on the generation of vehicles, pedestrians and bicycles into traffic than just the existing random process.
A DNN method will be developed for generating the predicition of the amount of people in the traffic. Historical traffic data and socio-economic mobility profiles created will be used for training the learning algorithm. The outcome will be a DNN algorithm that will have learned to predict the number of travelers of all transport profiles (vehicles, public transportation, pedestrian, bicyclist) passing the city area conditional on the weather, time of day, season, events and construction works at the area.
A Model Based DRL algorithm is developed into CARLA simulator to learn the best formation of the area outline and transportation modes using input provided by the generator developed. The most challenging part of the algorithm development is the design of the reward function (Lanzi 2002). Some research using RL for traffic flow prediction has been done, but their approach is very simplistic, e.g. they use very simple reward functions (Walraven et al. 2016). The reward function will be formed as a combination of smallest pollution effects, efficient commuting, livability, accessibility, and other factors agreed with stakeholders.