Predictive RTO for CoAP using Lightweight Support Vector Regression in Internet of Things

Sammanfattning

IoT is a rapidly expanding market that has introduced new challenges in the communication of constrained networks. Standard protocols such as CoAP, CoCoA, and CoCoA+ lack adaptability to dynamic network conditions, often resulting in inefficient congestion control in lossy or high-load IoT environments. These heuristic methods use different approaches for congestion control with predefined characteristics. Currently, ML models exist to improve congestion control; however, the area of lightweight protocols is not well established, as non-hybrid, full ML-based design methods are computationally heavy, which limits their practicality since IoT devices are often characterized by low memory and computational power. This work therefore aims to propose a lightweight model that can be implemented in resource-constrained IoT devices. The work investigates two questions: "How can a lightweight Support Vector Regression (SVR) model be developed to directly predict RTO values in CoAP networks, eliminating the need for traditional coefficient-based congestion control methods in IoT systems?" Furthermore, it aims to investigate to what extent this ML-based approach improves goodput and packet delivery ratio (PDR), and whether it remains within a reasonable amount of energy consumption. The simulation is built through a custom-developed simulator based on the standard protocol RFC 7252. The simulator generates data on which the model is later trained, making it a fully synthetic project. However, the simulators performance is validated against a real testbed. The two ML models used in this project are SVR for RTO prediction and Random Forest for the drop classifier. The ML models only use parameters known to the end-node device, making the model theoretically deployable in a real-world environment. The results report that the achieved simulation performance of the ML models competes with traditional protocols across all metrics. Furthermore, the project implements two SVR models, one more computationally complex and one linear, to indicate that computational overhead can be mitigated through rational ML design selection, favouring the simpler linear model. The linear model utilizes a per-attempt design for training, meaning it naturally enforces the logic of the traditional protocol CoAP. The results show that the linear model performs strongly and outperforms traditional protocols in terms of PDR. Although both ML models are outperformed in terms of goodput, the differences are marginal. The main limitation of this work is that the data is fully synthetic and constructed using a custom-built simulator rather than an established testbed such as Cooja or ns-3. Furthermore, while the linear per-attempt design achieves strong performance and the design works well, it can be vulnerable when real-world scenarios differ heavily from the training distribution.