HCRide uses multi-agent RL to boost ride-hailing fairness 5.4%

Ride-hailing dispatch algorithms typically optimize for operator revenue, often leaving passengers and drivers with unfair or unsatisfactory experiences. HCRide, a new system from researchers at (affiliations not specified in abstract), tackles this by explicitly modeling both passenger fairness and driver preference alongside system efficiency. At its core is Habic (Harmonization-oriented Actor-Bi-Critic), a multi-agent reinforcement learning algorithm with three components: a multi-agent competition mechanism, a dynamic Actor network, and a Bi-Critic network. This architecture lets the system navigate the inherent tension between giving passengers equitable service and respecting drivers' choices about which trips to accept.

The team validated HCRide on two real-world ride-hailing datasets from Shenzhen and New York City. Results show it beats state-of-the-art baselines by improving system efficiency 2.02%, passenger fairness 5.39%, and driver preference 10.21%. Notably, the gains are simultaneous—not traded off—suggesting it's possible to design dispatch systems that are both more human-centric and commercially viable. Accepted as a regular paper in the Human-Centred Artificial Intelligence Track at IJCAI 2025, HCRide points toward a future where ride-hailing platforms optimize for all stakeholders, not just the operator. The paper is available on arXiv (2508.04811).