COMNETS: COst-sensitive decision trees approach to throughput optimization for Multi-radio IoT NETworkS

Jothi Prasanna Shanmuga Sundaram, Magzhan Gabidolla, Miguel Á. Carreira-Perpiñán, Alberto E. Cerpa

Abstract

Mesoscale IoT applications, such as P2P energy trade and real-time industrial control systems, demand high throughput and low latency, with a secondary emphasis on energy efficiency as they rely on grid power or large-capacity batteries. MARS, a multi-radio architecture, leverages ML to instantaneously select the optimal radio for transmission, outperforming single-radio systems. However, MARS encounters a significant issue with cost sensitivity, where high-cost errors account for 40% throughput loss. Current cost-sensitive ML algorithms assign a misclassification cost for each class but not for each data sample. In MARS, each data sample has different costs, making it tedious to employ existing cost-sensitive ML algorithms. First, we address this issue by developing COMNETS, an ML-based radio selector using oblique trees optimized by Tree Alternating Optimization (TAO). TAO incorporates sample-specific misclassification costs to avert high-cost errors and achieves a 50% reduction in the decision tree size, making it more suitable for resource-constrained IoT devices. Second, we prove the stability property of TAO and leverage it to understand the critical factors affecting the radio-selection problem. Finally, our real-world evaluation of COMNETS at two different locations shows an average throughput gain of 20.83% and 17.39% over MARS.

Introduction

In the Internet of Things (IoT) landscape, the "mesoscale" (100–1000m) represents a challenging middle ground between short-range Bluetooth and long-range LoRa. Used for industrial automation and smart grids, these environments suffer from erratic link quality. To maintain robust connectivity, devices often employ multiple radios (e.g., Zigbee and LoRa), requiring a mechanism to select the optimal radio for every packet transmission.

The Challenge: High-Cost Selection Errors

Existing multi-radio systems like MARS utilize standard Machine Learning (ML) for radio selection. These models typically treat all errors equally using a "0/1 loss" approach. However, in radio selection, mistakes vary significantly in impact. A "low-cost" error occurs when the chosen radio is only slightly slower than the optimal one. Conversely, a "high-cost" error happens when the selected radio provides significantly lower throughput (e.g., 2000 bps vs. 7000 bps).

Figure 1: Comparison of radio selection errors and total throughput loss, illustrating how selection errors do not always correlate linearly with throughput loss.

Research shows that these high-cost errors can account for up to 40% of total throughput loss. Standard ML algorithms struggle here because they typically assign costs to classes rather than individual data samples, which is necessary given that IoT signal conditions change microsecond by microsecond.

The COMNETS Approach

COMNETS addresses these issues through three key technical innovations:

  1. Cost-Sensitive Tree Alternating Optimization (TAO): Unlike standard greedy decision trees (like CART), COMNETS uses the TAO algorithm. TAO allows the model to incorporate a specific "cost" for every training data point, directly tied to the throughput difference between radios at that moment.
  2. Oblique Decision Trees: While most trees look at one feature at a time (axis-aligned), COMNETS uses oblique trees. Each node evaluates a linear combination of multiple features simultaneously, resulting in a more powerful and efficient model.
  3. Sparsity and Stability: By applying $\ell_1$ regularization, the model is forced to use a minimal set of features (Sparse Oblique Trees). Furthermore, TAO-optimized trees exhibit structural stability, meaning the decision logic remains consistent even as new data is added, which is crucial for reliability in the field.

Experimental Setup and Datasets

The researchers validated COMNETS through a real-world deployment of a 15-node multi-radio mesh network using Raspberry Pi 3Bs equipped with Zigbee and LoRa radios.

Figure 6: The experimental hardware setup and deployment in real-world outdoor environments.

Testing was conducted at two distinct locations with varying building materials and human traffic. The team collected a dataset of 25,500 packets, recording features such as Hop Number (HN), RSSI (signal strength), and Required Number of Packets (RNP). Ground truth was established by manually labeling each packet with the radio that achieved the highest actual throughput.

Key Results and Insights

  • Throughput Gains: COMNETS outperformed MARS by 20.83% at Location A and 17.39% at Location B.
  • Efficiency: The resulting decision trees were 50% smaller than standard models, allowing for easy implementation as simple IF...ELSE logic on low-power microcontrollers.
  • Latency Management: In high-traffic scenarios, COMNETS maintained latency within the required 55ms bound, avoiding the queuing delays caused by poor radio choices in previous systems.
  • Interpretability: Sparse trees revealed that the model uses Hop Number for coarse distance-based decisions and RSSI/RNP for fine-grained local interference adjustments.

Conclusion

COMNETS demonstrates that a cost-sensitive approach to machine learning can significantly enhance the performance of multi-radio IoT networks. By prioritizing the avoidance of high-cost errors and utilizing efficient oblique trees, it provides a scalable and interpretable solution for high-throughput mesoscale applications. Future work will explore feeding raw packet sequences into the model and further optimizing implementation for ultra-low-power hardware.

Summary updated

Abstract

Mesoscale IoT applications, such as P2P energy trade and real-time industrial control systems, demand high throughput and low latency, with a secondary emphasis on energy efficiency as they rely on grid power or large-capacity batteries. MARS, a multi-radio architecture, leverages ML to instantaneously select the optimal radio for transmission, outperforming the single-radio systems. However, MARS encounters a significant issue with cost sensitivity, where high-cost errors account for 40% throughput loss. Current cost-sensitive ML algorithms assign a misclassification cost for each class but not for each data sample. In MARS, each data sample has different costs, making it tedious to employ existing cost-sensitive ML algorithms. First, we address this issue by developing COMNETS, an ML-based radio selector using oblique trees optimized by Tree Alternating Optimization (TAO). TAO incorporates sample-specific misclassification costs to avert high-cost errors and achieves a 50% reduction in the decision tree size, making it more suitable for resource-constrained IoT devices. Second, we prove the stability property of TAO and leverage it to understand the critical factors affecting the radio-selection problem. Finally, our real-world evaluation of COMNETS at two different locations shows an average throughput gain of 20.83%, 17.39% than MARS.
BibTeX
@inproceedings{sundaram2025comnets,
  author = {Jothi Prasanna Shanmuga Sundaram and Magzhan Gabidolla and Miguel Á. Carreira-Perpiñán and Alberto E. Cerpa},
  title = {{COMNETS: COst-sensitive decision trees approach to throughput optimization for Multi-radio IoT NETworkS}},
  booktitle = {ETFA},
  year = {2025},
  url = {https://openreview.net/forum?id=HHLbhiRPR0},
}
Figure