This repository illustrates a basic example of reinforcement learning on a traffic light system. For this, Keras is used on top of Tensorflow in conjunction with SUMO for the traffic simulation.
Main Dependencies ============
- Keras 2.2.2
- Numpy 1.14.5
- Tensorflow-gpu 1.10.0
- Gym 0.10.5
A number of python files are needed to train and test the desired deep reinforcement learning:
A collection of customized keras-rl objects for handling multimodel neural networks.
A couple of objects for deleting and generating new trips with random source/destination weights on each episode.
An environment class handling the simulation advancement, observation parsing and reward calculation.
Script for initiating the sumo environment and training the reinforcement learning model through Keras and Tensorflow.
sumoDqnTest.py ------------
Script for testing the saved model(s), by default through the sumo-gui.
Script for testing naive strategies, i.e. fixed phase-duration programs.