Abstract
Sparse Mobile Crowdsensing (MCS) aims to reduce data collection costs by sensing only a subset of cells and inferring the rest. This paper presents DR-Cell, a Deep Reinforcement Learning (DRL) framework for optimal cell selection. By employing a Deep Recurrent Q-Network (DRQN) with LSTM layers, DR-Cell captures spatial and temporal correlations to minimize sensing costs while meeting data quality constraints. Experimental results on real-world datasets show a 15% reduction in required samples compared to state-of-the-art heuristic methods.
Introduction
Mobile Crowdsensing (MCS) leverages smartphone sensors for urban monitoring while minimizing participant recruitment costs. The core challenge lies in selecting an optimal subset of "cells" (geographic sub-areas) to sense such that the remaining data can be accurately inferred. Traditional heuristic-based methods, such as Query-By-Committee (QBC), often fail to directly optimize for the minimum number of sensed cells under strict quality guarantees.
The sensing process involves iteratively choosing cells to collect data, which then informs the inference of unsensed regions.
As shown in the figure above, the system transitions through cycles, progressively filling the data matrix with collected and inferred values. This paper introduces DR-Cell, a DRL framework that treats cell selection as a sequential decision-making problem to maximize sensing efficiency.
The DR-Cell Framework
DR-Cell utilizes a Deep Recurrent Q-Network (DRQN) to learn an optimal sensing policy. The architecture maps the current state of the sensing grid to the most beneficial next action.
The framework consists of several key components:
- State Representation: The agent observes a "Cell Selection Matrix" representing the sensing history.
- Action Space: The agent selects the next cell to be sensed in the current cycle.
- Reward Mechanism: A reward function balances the cost of sensing against a significant bonus for achieving the target inference quality threshold.
- Recurrent Architecture: By incorporating Long Short-Term Memory (LSTM) layers, the DRQN captures both spatial dependencies between neighboring cells and temporal patterns over successive cycles.
To address data scarcity, the authors employ transfer learning. A model pre-trained on a data-rich "source" task (e.g., temperature sensing) is fine-tuned for a "target" task (e.g., humidity), leveraging shared spatial characteristics of environmental variables.
Experimental Evaluation
The authors evaluated DR-Cell on two real-world datasets: Sensor-Scope (environmental data from the EPFL campus) and U-Air (air quality data from Beijing). These datasets provide the ground truth necessary to simulate sparse sensing scenarios.
Results demonstrate that DR-Cell reduces the required number of sensed cells by up to 15% compared to QBC while maintaining identical inference quality. In temperature monitoring scenarios, DR-Cell maintained 90% quality using only ~22% of available cells. Furthermore, transfer learning significantly improved performance in scenarios with limited training data, highlighting the framework's generalizability.
Conclusion
DR-Cell represents a significant advancement in Sparse MCS by shifting from heuristic-based selection to an optimized, learning-based policy. By integrating recurrent neural networks with reinforcement learning, the system effectively exploits the spatio-temporal correlations inherent in urban data. Future work will focus on addressing the "cold start" problem through online learning and incorporating non-uniform sensing costs.
Summary updated