Computer Networks Laboratory

Historically, Service Function Chaining (SFC) optimization has been formulated as mixed-integer linear programs (MILP) and solved with heuristics. Those methods give interpretable, high-quality solutions but often cannot scale or meet online timing requirements. More recently, deep reinforcement learning (DRL) approaches have been explored to enable faster, online decision-making, though they typically reduce interpretability.

The cold-start problem is the delay and overhead when initializing on-demand software units (containers, serverless functions, or VNFs) in cloud environments. It matters most for systems that require rapid scale-out: startup latency can break strict QoS targets, so understanding and mitigating cold-starts is essential for practical deployments.

Indoor human detection aims to tell whether people are present behind walls or inside rooms using sensors placed outside or at a building’s perimeter. Traditionally this problem has relied on cameras, radar, or wideband RF—solutions that can be accurate but are expensive, intrusive, or hard to deploy at scale. Using low-frequency narrowband radios (e.g., ~433 MHz) trades raw precision for much better wall penetration, lower cost, and simpler hardware, but it also brings strong site-dependence, sensitivity to multipath and environmental change, and a need for careful thresholding — so the real question is whether such lightweight systems can be made robust and generalizable enough for dependable real-world use.