MULTI-AGENT REINFORCEMENT LEARNING WITH APPLICATION ON TRAFFIC FLOW CONTROL

Traffic congestion diminish driving experience and increases the CO2 emissions. With the rise of 5G and machine learning, the possibilities to reduce traffic congestion are endless. This thesis aims to study if multi-agent reinforcement learning speed recommendations on a vehicle level can reduce congestion and thus control traffic flow. This is done by simulating a highway with an obstacle on one side of the lanes, forcing all the vehicles to drive on the same lane past the obstacle, resulting in congestion. A game theory aspect of drivers not obeying the speed recommendations was implemented to further simulate real traffic. Three DeepQ-network based models were trained on the highway and the best model was tested. The tests showed that multi-agent reinforcement learning speed recommendations can reduce the congestion, measured in vehicle hours, up to 21% and if 1/3 of the vehicles uses the system, the total congestion can be significantly reduced. In addition, the test showed that the model achieves a success rate of 80%. Two improvements to the success rate would be more training and implementing a non reinforcement learning mechanism for the autonomous driving part