Dynamic-State-Traffic-Lights

This project presents a proposal to implement a dynamic state traffic light system, which adapts to conditions in real time to optimize the trafic flow.

This video shows a comparison between our trained model and a conventional traffic light system.

Instalation

In order to try our model, a basic installation of SUMO is needed (see SUMO oficial installation). Moreover, the Traffic Control Interface and SUMO python libraries must be installed to. You can install these libraries using pip:

pip install traci

pip install sumolib

Then, clone this repository:

git clone git@github.com:daniel-lima-lopez/Dynamic-State-Traffic-Lights.git

Move to the instalation folder

cd Dynamic-State-Traffic-Lights

Method description

The proposed method consists of a traffic light system whose states are controlled by a neural network, which receives as input the number of cars in each lane and predic the most next optimal state, as desribed in the next figure:

Regarding the neural network training, genetic algorithms are used to optimize the weights configuration. In this approach, the genetic algorithm seeks to minimize the time needed to handle a defined traffic flow. In this way, the algorithm optimizes the neural network configuration to control traffic appropriately.

Examples

To run the simulations presented in the video, you first need to move to Simulations folder:

cd Simulations

The file sim_base.py execute a simulation with a conventional traffic lights system. The file sim_nn.py execute a simulation with the same triffic configurations and the proposed method with a previously tranied neural networks, whose weight configuration is in the file optimo.txt.

To train the model from scratch, from the installation folder, move to Traning folder:

cd Training

The file genetico.py contains the traning routine which produces the optimal weight configuration (optimo.txt). It should be considered that the training process takes approximately 6 hours.