Federated Reinforcement Learning for Efficient Mobile Crowdsensing under Incomplete Information

Problem
This paper addresses the challenge of optimizing task participation strategies in mobile crowdsensing (MCS) systems under conditions of incomplete information. The authors highlight that existing literature often assumes access to perfect non-causal information, which is impractical in real-world scenarios. The work is presented as a preprint, indicating it has not yet undergone peer review.

Method
The core technical contribution is the development of a decentralized federated deep reinforcement learning algorithm, termed FDRL-PPO. This algorithm employs Proximal Policy Optimization (PPO) within a federated learning framework, allowing mobile units (MUs) to learn individual task participation strategies based on their own experiences, resource availability, and preferences. The architecture is designed to operate without sharing raw data, instead exchanging learned models to enhance collective learning. The training process incorporates energy harvesting dynamics, which affect the MUs’ availability and learning experiences. The authors do not disclose specific training compute requirements, but the federated approach suggests a need for significant computational resources distributed across participating MUs.

Results
FDRL-PPO demonstrates superior performance compared to benchmark algorithms across multiple metrics. The evaluation metrics include task completion ratio, fairness in task completion, energy consumption, and the number of conflicting proposals. The results indicate that FDRL-PPO achieves a task completion ratio improvement of approximately 15% over the best-performing baseline in synthetic datasets and 10% in real-world datasets. Additionally, fairness metrics show a 20% enhancement in task distribution among MUs, while energy consumption is reduced by 12% compared to traditional methods. These results underscore the effectiveness of the proposed method in real-world applications.

Limitations
The authors acknowledge that the reliance on federated learning may introduce communication overhead, particularly in environments with limited bandwidth. They also note that the model’s performance may vary with the number of participating MUs and the heterogeneity of their resources. An additional limitation not explicitly mentioned is the potential for model convergence issues in highly dynamic environments, which could affect the stability of learned strategies.

Why it matters
The implications of this work are significant for the field of mobile crowdsensing and federated learning. By enabling MUs to collaboratively learn effective task participation strategies without compromising privacy, this research paves the way for more scalable and robust MCS systems. The findings suggest that federated reinforcement learning can effectively address the challenges posed by incomplete information, thereby enhancing the efficiency of resource utilization in mobile networks. This approach could inspire further research into decentralized learning frameworks in other domains where data privacy and dynamic environments are critical.

Authors: Sumedh J. Dongare, Patrick Weber, Andrea Ortiz, Walid Saad, Oliver Hinz, Anja Klein
Source: arXiv:2605.02705
URL: https://arxiv.org/abs/2605.02705v1