Optimization of traffic management with learning machines
Résumé
This paper considers traffic management in metro networks, and techniques to allow fleets of trains to recover from bunching situations. We address this problem as a controller synthesis question. Controllers decide online train speeds and dwell durations to optimize the time needed to return to an ideal status of a network. We consider a formal model for metro lines based on a variant of Petri nets, and take as objective the minimization of the time needed to recover from a congestion. Though this is a standard timed game, the size of the state space for this model forbids standard synthesis techniques. We hence address the control question with ad-hoc local controllers, that take their decisions from neighbourhood of a train and ignore the rest of the network. We first propose a local controller that balances the headways w.r.t. the preceding and successor trains at each stop. The returned solution is obtained as the minimum of a quadratic cost function. We then replace the quadratic function minimization by a decision returned by a neural network trained on our metro model, and compare performances of both types of controllers on a case study.
Domaines
Théorie et langage formel [cs.FL]Origine | Fichiers produits par l'(les) auteur(s) |
---|