| 摘要: | 移動群眾感知系統（Mobile Crowd Sensing, MCS）近年已被廣泛視為收集大規模資料的方法，藉由眾多移動使用者的行動裝置（例如智慧型手機）內建感測器回傳城市環境資訊。當使用者投入群眾感知任務並回傳感測數據時，可以獲得相對應的獎勵，移動群眾感知平台則利用這些收集而來的資訊，設計與提供以使用者為中心的各項服務。
然而，移動使用者的高移動性與平台對其狀態認知的不完全，使得在招募使用者時同時兼顧任務完成率與成本控制格外困難。在此背景下，本研究將動態招募與激勵問題轉化為不完全資訊下的 Bayesian 二階段 tackelberg 賽局：平台先依信念對使用者移動模型作機率估計並公告資料單價，使用者再依自身移動行為與成本最佳回應，平台透過線上貝氏更新形成「學習–定價–回饋」的閉環決策流程，以兼顧預算控管與資料效益。
所提出之結合使用者移動性與賽局互動的 Bayesian Stackelberg 激勵方法，能根據互動結果動態調整資料單價，讓報酬與各類使用者的真實可用性及貢獻相符，並藉由信念更新自歷史資料學習使用者類型、優先招募高貢獻使用者。
本研究在單一 MCS 平台、多使用者多任務的模擬下驗證所提方法：在僅有 20% 積極使用者時，以約 500 的累積成本達成近 70% 任務完成率，成本僅為 GA-Stackelberg的約 60%，完成率僅低約 5%。當積極使用者提升至 80% 時，本方法用約 40% 預算即可達成約 39000 的平台效用，高於 RAIN 與 FCSMP 的約 27000，雖較 GA-Stackelberg效用低約 13%，但成本也減少約 35%，在預算與效用間取得良好平衡。進一步於Random Walk 與 San Francisco Taxi 真實軌跡及三種速度區間下，比較平台累積效用、激勵成本與任務完成率，結果顯示所提方法在不同移動模型與速度設定下皆能維持穩定表現，效能不受場景差異明顯影響。;Mobile Crowd Sensing (MCS) has in recent years been widely regarded as an effective paradigm for large-scale data collection, leveraging the embedded sensors of numerous mobile users＇devices (e.g., smartphones) to report urban environmental information. When users participate in crowdsensing tasks and upload sensing data, they receive corresponding rewards, while the MCS platform utilizes the collected information to design and provide various user-centric services.
However, the high mobility of users and the platform＇s incomplete knowledge of their states make it particularly challenging to simultaneously ensure task completion
and control costs during user recruitment. Against this background, this study reformulates the dynamic recruitment and rewarding problem as a Bayesian two-stage Stackelberg
game under incomplete information: the platform first performs probabilistic estimation of users＇mobility models based on its belief and announces the unit price of data; users
then best respond according to their mobility behavior and costs. The platform further conducts online Bayesian updates to form a “learning–pricing–feedback＂closed-loop decision process that balances budget control and data utility. The proposed Bayesian Stackelberg incentive mechanism, which integrates user mobility and game-theoretic interaction, allows the platform to dynamically adjust the unit data price according to interaction outcomes so that rewards are aligned with users＇true availability and contributions, and, through belief updates based on historical data, to learn user types and prioritize the recruitment of high-contribution users.
This study validates the proposed method in a simulated environment with a single MCS platform and multiple users and tasks. When only 20% of users are active, the
method achieves nearly a 70% task completion rate with a cumulative cost of about
500, which is roughly 60% of the cost of GA-Stackelberg, while the completion rate
is only about 5% lower. When the proportion of active users increases to 80%, the proposed method attains a platform utility of about 39,000 using only around 40% of
the total budget, higher than the approximately 27,000 utility achieved by RAIN and FCSMP. Although its utility is about 13% lower than that of GA-Stackelberg, its cost
is also reduced by about 35%, yielding a favorable balance between budget and utility.
Furthermore, under the Random Walk model and real trajectories from the San Francisco Taxi dataset, as well as three different speed ranges, we compare cumulative platform
utility, incentive cost, and task completion rate. The results show that the proposed method maintains stable performance across different mobility models and speed settings, with its effectiveness not significantly affected by scenario variations. |
