Optimal Chaining of Vehicle Plans with Time Windows

For solving problems from the domain of Mobility-on-Demand (MoD), we often need to connect vehicle plans into plans spanning longer time, a process we call plan chaining. As we show in this work, chaining of the plans can be used to reduce the size of MoD providers' fleet (fleet-sizing problem) but also to reduce the total driven distance by providing high-quality vehicle dispatching solutions in MoD systems. Recently, a solution that uses this principle has been proposed to solve the fleet-sizing problem. The method does not consider the time flexibility of the plans. Instead, plans are fixed in time and cannot be delayed. However, time flexibility is an essential property of all vehicle problems with time windows. This work presents a new plan chaining formulation that considers delays as allowed by the time windows and a solution method for solving it. Moreover, we prove that the proposed plan chaining method is optimal, and we analyze its complexity. Finally, we list some practical applications and perform a demonstration for one of them: a new heuristic vehicle dispatching method for solving the static dial-a-ride problem. The demonstration results show that our proposed method provides a better solution than the two heuristic baselines for the majority of instances that cannot be solved optimally. At the same time, our method does not have the largest computational time requirements compared to the baselines. Therefore, we conclude that the proposed optimal chaining method provides not only theoretically sound results but is also practically applicable.

Stochastic Langevin Differential Inclusions with Applications to Machine Learning

Stochastic differential equations of Langevin-diffusion form have received significant recent, thanks to their foundational role in both Bayesian sampling algorithms and optimization in machine learning. In the latter, they serve as a conceptual model of the stochastic gradient flow in training over-parametrized models. However, the literature typically assumes smoothness of the potential, whose gradient is the drift term. Nevertheless, there are many problems, for which the potential function is not continuously differentiable, and hence the drift is not Lipschitz-continuous everywhere. This is exemplified by robust losses and Rectified Linear Units in regression problems. In this paper, we show some foundational results regarding the flow and asymptotic properties of Langevin-type Stochastic Differential Inclusions under assumptions appropriate to the machine-learning settings. In particular, we show strong existence of the solution, as well as asymptotic minimization of the canonical Free Energy Functional.