RL for Edge Resource Scaling: Can Federated Metadata Replace Centralized Control?

[rishabkh]

Hello Community!

As edge computing grows, dynamic workloads and resource constraints (e.g., intermittent connectivity, thermal throttling) challenge traditional centralized resource managers. These systems often struggle with latency, scalability, and single points of failure.

Idea: What if we replaced centralized control with federated metadata-driven reinforcement learning (RL)? Instead of relying on a global orchestrator, edge nodes could share lightweight telemetry (e.g., gradient staleness, accelerator utilization) to train lightweight RL agents locally. This could enable:

1. Adaptive scaling: Nodes autonomously adjust resources using real-time metadata.
2. Reduced overhead: Eliminate centralized coordination bottlenecks.
3. Fault tolerance: Decentralized decisions prevent system-wide failures.

Questions for Discussion:

1. Have you encountered edge scaling problems where centralized approaches fell short?
2. What metadata (e.g., hardware telemetry, model updates) would be most valuable for RL agents?

Could this complement existing frameworks like Kubernetes or ROS?

Background: I’m exploring this for my thesis (building on prior RL work for edge-cloud systems) and would love your insights! Let’s brainstorm how federated learning and edge ML systems can co-evolve.

Best,
Rishab Khanna (BITS Pilani | Ex- Harvard Research Affiliate)