CityFlowER: Efficient and Realistic Traffic Simulation

CityFlowER is An Efficient and Realistic Traffic Simulator Platform with Embedded Machine Learning Models. It Can Support Multi-agent Reinforcement Learning for Various Traffic Analysis and Control Tasks.

Framework and comparison to other implementations:

This simulator introduces following features:

• Realistic A microscopic traffic simulator which simulates the behavior of each vehicle. Any individual vehicle can be controlled by one or multiple behavior models to mimic the realistic driving scenario.

• Efficient The embedded structure's simulation speed is much faster than API based implementation, and approximating the rule based control methods.

• Large Scale CityFlowER support large scale simulation efficiently.

• Simple config setting for loaded models. You could find the following implementation in src/vehicle/vehicle.cpp or src/vehicle/lanechange.cpp The code blocks to import rule based file is shown below: We also provide training example in Imitate_cityflow/*, please check it out for example of training:

  // vehicleInfo is updated out of the function
  // "PATH" is the pre-trained model path
  double Vehicle::getCarFollowSpeed(double interval) {
  
  Vehicle *leader = getLeader();
  std::vector<torch::jit::IValue> input_feature;
  auto input = torch::tensor ({leader->getSpeed(), vehicleInfo.speed ...});
  
  input_feature.push_back(input);
  static torch::jit::script::Module speedModel = torch::jit::load("PATH");
  
  at::Tensor output_speed = speedModel.forward(input_feature).toTensor();
  return output_speed.item<double>();
  }

  void SimpleLaneChange::makeSignal(double interval) {
  
  std::vector<torch::jit::IValue> input_feature;
  auto input = torch::tensor ({vehicle->engine->getCurrentTime(),...});
  input_feature.push_back(input);
  
  static torch::jit::script::Module laneModel = torch::jit::load("PATH");
  at::Tensor lane_choice = laneModel.forward(input_feature).toTensor();
  signalSend->target = lane_choice;
  }

  For large scale vehicle settings, you could also define control_config.xml and mapping the vehicle_id with assigned ML models intended to use.

  In next updated version, we will provide this detailed instructions for the defination of control_config.xml.

Release Plans

Release Checklist

• Basic Platform Code
• Integration with LibTorch
• Functionality Release
  • Car Following Speed Model
  • Lane Changing Model
• Install & Deployment guideline
• Pre-trained model release

Featured Research and Projects Using CityFlow

• PressLight: Learning Max Pressure Control to Coordinate Traffic Signals in Arterial Network (KDD 2019)
• CoLight: Learning Network-level Cooperation for Traffic Signal Control
• Traffic Signal Control Benchmark
• TSCC2050: A Traffic Signal Control Game by Tianrang Intelligence (in Chinese) Tianrang Intelligence home page

Links

• Home Page
• Simulation Demo Video