TSDM: A Scheduling Policy for Joint Throughput-AoI Optimization in Multichannel Wireless Networks

What the math gives to ML

The paper offers a second-order scheduling mechanism: instead of optimizing only average service or throughput, it first assigns each flow a target mean delivery rate and temporal variance, then uses a deficit-based weighted matching policy to track those targets online. The transferable asset is a low-complexity controller for allocating scarce update opportunities among competing neural-network components while controlling both average progress and burstiness. A direct neural analogue is to schedule parameter blocks, task heads, experts, or replay sources according to deficits in their desired update frequency and update-frequency variance. This is most promising for multi-task training, mixture-of-experts load balancing, or asynchronous distributed optimization, where starvation and bursty updates can destabilize learning.