We are developing models to analyze and optimize the cost of transit operations by focusing on the energy impact of the vehicles. For this purpose, we are developing real-time data sets containing information about engine telemetry, including engine speed, GPS position, fuel usage, and state of charge (electrical vehicles) from all vehicles in addition to traffic congestion, current events in the city, and the braking and acceleration patterns. These high-dimensional datasets allow us to train accurate data-driven predictors using deep neural networks, for energy consumption given various routes and schedules. Having these predictors combined with traffic congestion information obtained from external sources will enable the agencies to identify and mitigate energy efficiency bottlenecks within each specific mode of operation such as electric bus and electric car. To make this possible, the project is also developing new distributed computing and machine learning algorithms that can handle data at such a rate and scale.

We are developing microtransit dispatch algorithms that can serve passengers using dynamically generated routes and may expect passengers to make their way to and from common pick-up or drop-off points. Our hypothesis in the project is that the integration of data-driven methods with better operational research methods combined with a socially engaged design will lead to success. A key aspect of this research area is the development of techniques to preserve privacy across multimodal datasets, while also providing sufficient information for analysis and scheduling. This approach is different from commercial alternatives, because most commercial alternatives emphasize only on economic objectives, and as such, those services are often not integrated with public transit and do not address equity issues, which is a critical concern for us. The outcome of the project will be a deployment-ready software system that can be used to design and operate a micro-transit service effectively.

Lastly, we are developing algorithms to perform system-wide optimization, (the microtransit, fixed line and paratransit) focusing on three objectives: minimizing energy per passenger per mile, minimizing total energy consumed, and maximizing the percentage of daily trips served by public transit. While it is possible to optimize these decisions separately as prior work has done, integrated optimization can lead to significantly better service (e.g., synchronizing flexible courtesy stops with microtransit dispatch for easy transfer). However, this is hard due to uncertainty of future demand, traffic conditions etc. We address these challenges using state-of-the-art artificial intelligence, machine learning, and data-driven optimization techniques. Deep reinforcement learning (DRL) and Monte-Carlo tree search form the core of our operational optimization, which is supported by data-driven optimization for offline planning and by machine learning techniques for predicting demand, maintenance requirements, and traffic conditions.