Multi-Agent Interactive Localization: A Positive Transfer Learning Perspective

Sharing and transferring knowledge among multiple targets (multi-agent) can improve joint environmental cognitive capability. While it is often overlooked in indoor localization, which is a typical application scenario of cognitive networks. In this paper, we propose a framework called Multi-Agent Interactive Localization (MAIL) to enhance overall localization performance. The framework consists of constructing a private and personalized localization model for each agent, encoding environmental knowledge in the model parameters, and sharing the parameters in a multi-agent cloud for subsequent interactions. We introduce a positive and privacy-preserving transfer learning (3PTL) model to transfer parameter-based knowledge among agents. Moreover, we explain the essence of negative transfer learning (NTL), when it can occur, and how to select transfer parameters reasonably to avoid NTL through rigorous theoretical analysis. Hence, the proposed 3PTL scheme unified in MAIL can simultaneously improve localization performance and ensure information privacy for multi-agent systems even with limited knowledge. Finally, we conduct extensive real-world experiments and simulations to validate the effectiveness of our proposed theoretical results and the superiority of the proposed 3PTL scheme.