Abstract

Federated learning enables collaborative model training among distributed devices without sharing raw data, making it promising for intelligent traffic monitoring systems. However, static client-selection policies fail under the dynamic, non-IID, and energy-limited conditions typical of urban traffic environments. This paper proposes a federated meta-Bayesian reinforcement learning framework that integrates Bayesian meta-learning with reinforcement scheduling to adaptively select clients in traffic intelligence applications. The system updates hierarchical priors from distributed traffic data streams and optimizes client participation policies using a multi-objective reward function balancing prediction accuracy, energy efficiency, and latency constraints. In our traffic-adapted experimental setup using non-IID MNIST data (simulating heterogeneous traffic patterns with two shards per client), The proposed framework achieves a mean accuracy of 0.84 with an 80% reduction in resource costs compared to conventional federated learning. The method provides an adaptive, energy-efficient, and self-optimizing client scheduling mechanism specifically designed for real-time federated traffic monitoring systems.

Statistics