Dynamic Lane Reversal: A reinforcement learning approach

The concept of dynamic space allocation has attracted increasing interest lately due to the advent of connectivity and autonomous era. Modern cities are trying to transcend the limits of “traditional’’ static urban design and utilize systems that dynamically allocate street space according to the prevailing conditions. The development of Dynamic Lane Reversal (DLR) schemes is one of the most popular ways being investigated towards this goal. Most DLR schemes found in the literature are computationally intensive and hard to implement in large scale real-life scenarios and ignore the microscopic idiosyncrasies of lane reversals. In this paper, a reinforcement learning scheme is developed that learns when to perform a lane reversal on an artificial intersection. The algorithm is trained in a base scenario within SUMO (Simulation of Urban MObility) microscopic simulation and then tested in different demand patterns, proving its excellent transferability. Finally, the results indicate that the proposed scheme leads to reducing the average trip duration (up to 19.9% - depending on the demand scenario) and the greenhouse gas emissions of the intersection (up to 15.10% for CO2).