Reinforcement Learning in Traffic Control for Connected Automated Vehicles

Abstract

The last years, more and people are concentrating in big cities for reasons of living and working. This effect has already some negative impacts on transportation networks including congestion and inefficiency. Parallel to the centralization, the number of autonomous vehicles on roads is continuing to grow, without completely replacing human driving vehicles. The upcoming mixed autonomy traffic situations will bring more dangers in terms of safety and transportation efficiency. The traditional traffic management solutions may not be able to handle these situations. Machine learning approaches have been already proved efficient in various complex fields. In this dissertation, a sub-field of Machine Learning, the Deep Reinforcement Learning will be investigated for enabling a smooth coexistence of automated, connected, and conventional vehicles. In particular, various reinforcement learning models, with both single and multi agent approaches, will be trained and tested on controlling the traffic flow in a specific mixed autonomy traffic scenario, where a transition from autonomous to human driving mode is needed for the vehicles.