TFTC Webinars: Dr. Sabya Mishra -- Thursday, February 10th @ 3 PM EST
Rafegh Aghamohammadi
Title: Optimal Rebalancing and Matching Strategy for Autonomous and Human Driven Taxi Fleet using Reinforcement Learning
Speaker: Dr. Sabya Mishra, University of Memphis
Time: Thursday, February 10th, 3:00-4:30 PM EST
Meeting Link: https://umich.zoom.us/j/92938723826 (Passcode: NGTS)
Abstract:
With the advancement of technology, transportation network companies (TNCs) are able to provide better ride-hailing service compared to the traditional taxi company. However, this service is still hampered by the imbalance between supply and demand in both spatial and temporal dimensions. Surge pricing appears as a current solution, but its effectiveness is questionable. Prior studies found that riders are more likely to cancel trips due to excessive price and thus, drivers lose potential customers. Another potential solution is to introduce Autonomous Vehicles (AV) Taxi, which is likely to be electric, into the current fleet and strategically rebalance these AVs from low to high demand areas. However, in the near future integration of AV, TNCs is still likely to consider human-driven vehicles (HV) taxi within their system. This adds another layer of complexity since TNCs need to avoid disappointing HVs by making them wait for too long until a match with customer appears. In addition to rebalancing, single-pickup matching for both HV and AV, TNCs also consider a potential carpooling and charging of AV action. This research proposes a framework for matching (single-pickup and carpooling), rebalancing, and recharging of AV while simultaneously satisfying HV taxis and minimizing riders’ cost such as wait time and cancellation. The single-pickup and carpooling processes are modeled as a Mixed Integer Programming whereas the rebalancing and recharging action are formulated as a Reinforcement Learning (RL) model. The goal of RL is to find the optimal policy function by considering inputs such as demand and supply; and determine how many vehicles are needed to rebalance from a specific zone to another and how many to go charging. The RL model is trained by an asynchronous solving algorithm which enables the paper’s framework to be computationally capable with real-world applications. A case study and demonstration of the framework is performed on the Anaheim, California network with simulated taxi demand and supply. The results indicated that the framework reduces the waiting time and cancellation by 75% and 72% respectively compared to the baseline scenario. This research is beneficial to not only private TNC but also the riders with considerably improved riding experiences.
Traffic flow webinars
Title: Optimal Rebalancing and Matching Strategy for Autonomous and Human Driven Taxi Fleet using Reinforcement Learning
Speaker: Dr. Sabya Mishra, University of Memphis