Hybrid State Value Networks for Ride-Hailing Dispatch

The patent combines offline-learned dispatching policies with online-learned policies into a hybrid framework. The offline component provides robustness through historical data training, while the online component adapts to real-time changes through continuous learning from current system states and rewards. This merging allows the system to simultaneously achieve both robustness and responsiveness.

The dispatching policy transitions from a static offline-learned approach to a dynamic hybrid approach that continuously adapts. The online learning component updates policy parameters in real-time based on current system states and observed rewards, enabling the policy to dynamically adjust to changing conditions while maintaining the stability provided by offline learning.

The patent merges online learning with offline learning to balance responsiveness and robustness. The offline-learned policy serves as a stable foundation trained on comprehensive historical data, while the online component provides real-time adaptability. This combination prevents the system from being overly responsive to noisy real-time signals while maintaining the ability to adapt to genuine changes.

The offline-learned policy acts as a cushion or buffer against the variability and potential noise in online learning. By pre-training on historical data, the system establishes a robust baseline that cushions against over-reactivity to short-term fluctuations, while still allowing the online component to provide necessary real-time adaptations.