More about HKUST
Order Dispatching for Electric Vehicles by Deep Multi-Objective Reinforcement Learning
MPhil Thesis Defence Title: "Order Dispatching for Electric Vehicles by Deep Multi-Objective Reinforcement Learning" By Mr. Jinliang DENG Abstract The electrification of fleets is intensifying in on demand ride-hailing platforms (e.g. Uber, Lyft, and Didi Chuxing), but a major technical barrier for deploying electric vehicles is that an electric vehicle has a long charging time and relatively lower mileage when fully charged than fully filled conventional fuel vehicles. Subsequently, a major challenge for the ride-hailing platform is how to adopt the unique characteristics of electric vehicles into their order dispatching system. The aim of this work is to develop a formal dispatching framework that targets at improving the incoming of electric vehicle drivers on the platform, while considering attainability to charging stations to each vehicle. Specifically, we formulate the dispatching problem as a Mutli-Objective Markov Decision Process (MOMDP) and solve it with a novel reinforcement learning approach. Our approach that generates the candidate driver-order matching pairs consists of two steps: (i) the estimation of potential gain for each possible matching; and (ii) the elimination of the matching pairs that lead the electric vehicles to an out-of-power situation. The candidate matching pairs with the corresponding gain are then passed through the Maxflow algorithm to extract a consistent combination of pairs with the highest aggregated gain. The experiments show that our algorithm can improve the driver income, while preventing electric vehicles from running out of power on the road. Date: Monday, 29 April 2019 Time: 4:00pm - 6:00pm Venue: Room 4621 Lifts 31/32 Committee Members: Prof. Qiang Yang (Supervisor) Dr. Xiaojuan Ma (Supervisor) Prof. Fangzhen Lin (Chairperson) Dr. Yangqiu Song **** ALL are Welcome ****