Negotiation-augmented federated reinforcement learning for conflict-free edge–cloud stream scheduling

The rapid expansion of latency-sensitive streaming applications in edge-cloud systems presents considerable hurdles for adaptive scheduling, especially amid variable workloads, diverse resources, and multi-cluster interactions. Current decentralized methodologies, such as Federated Reinforcement Learning (FRL), facilitate privacy-preserving decision-making; however, they frequently encounter policy conflicts, uncoordinated adaptations, and restricted global coherence, resulting in resource contention, excessive task migrations, and diminished Quality of Service (QoS). To overcome these limitations, we propose FedNeg-RL (Federated Negotiation-Augmented Reinforcement Learning), a multi-agent adaptive scheduling framework that combines FRL with a streamlined inter-agent negotiation mechanism to ensure conflict-free and globally consistent decisions in distributed edge-cloud systems. In FedNeg-RL, cluster-level agents acquire local scheduling policies while collaborating through a negotiation layer based on structured argumentation protocols, facilitating proactive settlement of inter-cluster disputes without the exchange of raw data. The methodology additionally integrates workload forecasting and negotiation-aware clustering to predict demand fluctuations and synchronize scheduling activities across clusters. Comprehensive simulations utilizing Internet-of-Things (IoT)-driven streaming workloads reveal that FedNeg-RL diminishes conflict-induced reconfigurations by as much as 41%, reduces 90th-percentile latency by 20-28%, and lessens adaptation overhead by roughly 35% in comparison to leading FRL-based benchmarks. These findings validate that FedNeg-RL markedly enhances coordination, stability, and quality of service in highly dynamic edge-cloud systems.