RL-based anticipatory routing and matching in on-demand ridepooling

Dynamic ridepooling systems must constantly face a trade-off between current service quality and anticipated future demand. This trade-off is often addressed by rebalancing idle vehicles, with few methods incorporating anticipation into routing and matching decisions. Building upon previous research, where the weight of the anticipatory component—represented by a parameter θ—is manually preset and fixed throughout, we identify a key limitation: such static settings may fail to adapt to evolving system conditions. To address this, we propose a method in which θ is computed endogenously and dynamically adjusts in response to environmental changes, encouraging vehicles to end their trips in high-demand zones as appropriate. To achieve this, we propose a reinforcement learning framework that integrates a Graph Neural Network (GNN) with an Actor-Critic architecture, enabling the model to learn spatiotemporal demand-supply patterns and generate policies that balance short-term and long-term efficiency. Through simulations on real-world data from Manhattan (USA) and Utrecht (Netherlands), our approach consistently outperforms fixed-parameter and non-anticipatory baselines by reducing both rejection rates and user costs. Notably, we observe a seed-and-harvest pattern in the learned policy, where early anticipation improves downstream performance. The proposed method generalizes well across cities, highlighting its robustness and potential for deployment in diverse urban mobility systems.