Self-Interested Coalitional Crowdsensing for Multi-Agent Interactive Environment Monitoring.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Mobile crowdsensing (MCS) leverages distributed sensing capabilities for large-scale services like smart transportation and environmental monitoring.
• Multi-agent reinforcement learning (MARL) strategy training demands extensive environment interaction, leading to high costs.
• Complex sensing environments generate sparse, heterogeneous data, hindering accurate environment reconstruction.

Purpose of the Study:

• To develop a robust multi-agent environment monitoring framework (SCC-MIE) addressing challenges in data sparsity and heterogeneity.
• To improve the accuracy and efficiency of environment reconstruction and worker selection in MCS.
• To reduce the costs associated with MARL strategy training in MCS.

Main Methods:

• Developed a self-interested coalitional learning strategy within a multi-agent generative adversarial imitation learning framework.
• Integrated a reconstructor and discriminator for collaborative learning of the sensing environment and hidden confounders.
• Employed the secretary problem for real-time selection of optimal workers for data collection.

Main Results:

• SCC-MIE framework demonstrated significant performance improvements in environment monitoring compared to existing models.
• The approach effectively handles sparse and heterogeneous data for more accurate environment reconstruction.
• Achieved enhanced interpretability in environment monitoring results through cooperative learning.

Conclusions:

• SCC-MIE offers a robust and cost-effective solution for multi-agent environment monitoring using MCS.
• The proposed self-interested coalitional learning strategy enhances cooperation and learning accuracy.
• This framework provides a promising direction for advanced applications in smart environments and data-driven services.