Reinforcement Learning Augmented Optimization for Smart Mobility

Many mobility applications in smart cities are addressed as optimization problems. However, often, these problems are fragile due to their large-scale and non-convex nature, and also due to uncertainties arising because of human activity. In this paper, we apply a model-based Markov-decision-process (MDP) closed-loop identification algorithm to augment classical optimizers, with a view to alleviating this fragility. Specifically, we use deterministic optimal solutions provided by classical optimizers as initial guesses for MDP's policies, which are then "amended" as a result of online interaction with the environment to cope with uncertainty. Applications are described from niche of smart mobility problems, and numerical results are provided.