Dynamic Public Resource Allocation Based on Human Mobility Prediction

In dense points‑of‑interest (POI) areas such as amusement parks, static deployment of resources like trash bins leads to high costs and low utilization because crowd distribution varies greatly over time and space. Observations at Beijing Happy Valley show that 70‑80% of the 95 installed bins are under‑used.

The paper asks whether mobile resources can achieve the same coverage with fewer units and proposes a dynamic deployment framework based on human mobility prediction.

Three main challenges are identified: limited resource budget, energy constraints of mobile agents, and the large action space for scheduling.

To address the first two challenges, a multi‑step crowd flow prediction is performed, and a Multi‑Agent Long‑term Maximal Coverage Scheduling (MALMCS) problem is formulated, which seeks position sequences for agents that maximize total crowd coverage without exhausting energy.

Because MALMCS is NP‑hard, a two‑stage heuristic called Energy Adaptive Scheduling (EADS) is introduced. The first stage solves a multi‑level maximum‑k‑coverage problem using a greedy algorithm and a bottleneck assignment approach; the second stage applies an energy‑aware hill‑climbing adjustment when the greedy solution violates energy limits.

Model Predictive Control (MPC) is employed to re‑solve MALMCS at each decision point using the latest crowd flow forecasts, mitigating long‑term prediction errors.

Experiments use ten months of hourly crowd‑flow heat‑maps from Tencent location data (5,508 grid cells) and the real positions of 95 trash bins. Evaluation metric is the average daily covered crowd index (ADCC). Results show that the EADS‑MPC approach achieves comparable coverage with only 19 mobile resources, saving roughly 80% of static resources compared to the current deployment and 56.8% compared to a static maximal‑k‑coverage baseline.

References to related work on spatio‑temporal neural networks, budgeted maximum coverage, and multi‑level bottleneck assignment are provided.