Shared Mobility On Demand System Design

Tomorrows mobility will be radically different. Connected, Autonomous, Shared , and Electric Mobility are four main developments that are dramatically altering the automobile industry. We study the shared centralized class of mobility problems which considers a platform of self driving cars. There are new challenges with these systems such as how to balance the idle vehicle, how to price the shared autonomous system, and etc. We are attempting to address the question of how to share passengers ride to maximize satisfaction for riders, and the platform itself. Besides that, to have a good ETA estimate for trips, we develop a data-driven travel time prediction algorithm which can be used in our platform to get a good estimate for scheduling and routing the rides. Finally, we also study the pricing mechanism of these systems using a deep reinforcement learning agent that simulates the rides in New York. We start by studying both static and dynamic (real-time) ride pooling problem with time windows, multiple homogeneous/heterogeneous vehicles, passenger convenience and other business considerations. First, the problems under consideration is modeled as two different static MILP for homogeneous/heterogeneous fleet of vehicles, and also a constraint programming counterpart is provided for the heterogeneous vehicles case. Also to improve the linear relaxation of these models, several pre-processing steps and lifting inequalities are applied. While appealing, exact formulations have integer variables which render them as non-convex optimization problems. Thus, while this approach offers the benefit of system optimality, its formulation here is NP-hard, making it not viable for real world problems. To find a good quality solution, a heuristic decomposition algorithm based on constraint programming and branch and price is proposed to solve static model within a reasonable time for implementation in a real-world situation. Computational results show that the heuristic algorithms are superior compared to the exact algorithms in terms of the calculation time as the problem size (in terms of the number of requests) increases. In phase 2 of this dissertation, we propose a travel time predictive model by developing a integrated multi-step approach to learn the feature space. This multi stage algorithm is initiated by pre-processing task. Subsequently, the feature set is obtained by incorporating some publicly available information. Moreover, a feature engineer ing path is proposed to improve the feature space. This path includes Principal Component Analysis (PCA), geospatial features analysis, and unsupervised learning methods like K-Means and stacked autoencoders. Finally we apply a customized gradient boosting method to estimate travel times and comparing our results with LSTM network which shows superiority of our method in terms of capturing dynamics of traffic through time. Although more data with rare events need to be added in case of experiencing heavy snow or other events which magnifies travel times. Lastly, we developed a fleet management simulation platform where we model pricing problem as a partially observable Markov decision process (POMDP), and DQN agent is developed to estimate fares as a function of real-time interaction with the environment. Fare prices are considered to be continuous and stochastic variables, but for simplicity we have price adjustment in discrete units, and we determine them using a deep neural network (DNN). We compare our algorithm with the one for ride hailing system and see if our pricing mechanism can decrease rejections and cancellation and increase system objective as well as passengers’ utility. We illustrate the usefulness of our algorithm by applying it to real-world transportation problem and show that it learns fare estimates to minimize total travel time, maximize revenue, and other weighted objectives. Collectively, this work can be used for designing a ride sharing system of autonomous vehicles in which a controller module with many different predictive and prescriptive analytics engines dispatches vehicles and broadcasts ride fares to optimize system and riders utility