Title:Near Optimal Control of a Ride-Hailing Platform via Mirror Backpressure

Abstract:Ride-hailing platforms need to match supply and demand so as to maximize the rate of payoff generation while respecting the geographical flow constraints, namely, that the rate at which vehicles arrive at a location must equal the rate at which they leave that location. The platform's control levers include: (i) entry control, i.e. it can choose not to serve some customers, (ii) dynamic pricing, i.e. it can set prices for each ride depending on its origin and destination, and (iii) assignment rule, i.e. it can choose from which neighboring location to dispatch a car for pickup.
We consider two settings depending on whether pricing is an available lever to the platform. In joint-entry-assignment (JEA) setting, the platform can only use levers (i) and (iii) (for example, Didi in China does not use dynamic pricing). In joint-pricing-assignment (JPA) setting, the platform can use levers (ii) and (iii) (e.g. most ride-hailing platforms in North America). We introduce a novel family of Mirror Backpressure (MBP) platform control policies which are simple, practical, and do not require prior knowledge of the demand arrival rates. Key challenges include that serving a customer reduces supply availability at the dispatch location but increases it at the dropoff location ("supply externalities"), and that the number of vehicles at any location can never be negative (no-underflow constraints"). Mirror Backpressure generalizes the backpressure policy such that it executes mirror descent, allowing us to address these challenges. MBP loses at most an $O(K/T + 1/K)$ fraction of the achievable payoff in the JEA setting and $O(\sqrt{K/T + 1/K})$ fraction in the JPA setting, where $K$ is the number of vehicles and $T$ is the horizon. Simulation results in a realistic environment support our theoretical findings.