prCoAP uses lightweight SVR to boost IoT packet delivery to 97.25%
A new AI-driven CoAP variant achieves 97.25% PDR on low-end microcontrollers.
Internet of Things (IoT) networks rely on CoAP (Constrained Application Protocol) for lightweight messaging, but its congestion control still uses fixed heuristics like binary exponential backoff (BEB) and RTT-based mechanisms (CoCoA, CoCoA+). These do not adapt well to dynamic, lossy wireless links. To solve this, Tobias Hansson and Praveen Kumar Donta introduce prCoAP, a data-driven replacement for Retransmission Timeout (RTO) selection. prCoAP uses a per-attempt linear Support Vector Regression (SVR) ensemble that predicts RTO directly from node-observable features, running on-device on low-end microcontrollers with strict memory and energy budgets. A calibrated Random Forest drop classifier further identifies transactions likely to fail in later retransmissions and terminates them early, reducing channel occupancy.
The approach was evaluated using a discrete-event simulator implementing IEEE 802.15.4 and RFC 7252, and validated with the FIT IoT-LAB testbed. The linear SVR achieved 97.25% packet delivery ratio (PDR), outperforming standard CoAP under the evaluated conditions. A kernel SVR variant improved regression fit (R² 0.84 vs. 0.63) but the linear SVR provided better system-level efficiency, delivering comparable PDR with lower energy overhead. prCoAP demonstrates that lightweight machine learning can significantly improve IoT reliability without exceeding constraints of memory, compute, or power.
- prCoAP replaces heuristic RTO with a linear SVR ensemble, achieving 97.25% PDR on constrained IoT devices.
- A Random Forest drop classifier terminates likely-failed retransmissions early, reducing channel occupancy.
- Linear SVR outperforms kernel SVR in system efficiency despite lower R² (0.63 vs 0.84).
Why It Matters
prCoAP brings adaptive AI congestion control to low-power IoT, improving reliability on lossy links without hardware upgrades.