Efficient utilisation of urban road networks has been in the epicentre of researchers’ attention for many decades already. Nowadays, many cities rely on traffic-responsive strategies for the effective management of traffic flows. However, state of the art methodologies, based on Artificial Intelligence (AI), have started challenging the currently implemented solutions. Reinforcement Learning (RL) stands as one of the most promising AI-based methodologies aiming at the optimization of traffic signal controlling. Since the first appearance of relevant RL methodologies, numerous implementation strategies have been suggested. Nonetheless, these approaches are usually evaluated under ideal or simplistic conditions (e.g. toy networks, unrealistic demand patterns, etc.) thus lack the ability to assess the effectiveness of RL-based signalling optimisation under realistic conditions. The currently presented study bridges this gap by evaluating a plethora of RL-based implementations under fully realistic urban demand conditions as manifested in the large-scale road network of Nicosia, Cyprus.

Download: (a) Extended Abstract (b) Presentation video