COMNETS: COst-sensitive learning for throughput optimization in Multi-radio IoT NETworkS

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

In the world of the Internet of Things (IoT), we often talk about "short-range" (like Bluetooth in your home) or "long-range" (like LoRa across a city). But there is a tricky middle ground—the mesoscale (100–1000m)—used for industrial automation and smart grids. In these environments, link quality is erratic. To stay connected, devices often use multiple radios (e.g., Zigbee and LoRa) and try to pick the best one for every packet.

The new paper COMNETS introduces a cost-sensitive machine learning approach that doesn't just try to be "accurate"—it tries to avoid the specific mistakes that hurt throughput the most.

TL;DR

COMNETS is a radio-selection framework for multi-radio IoT networks that uses cost-sensitive oblique decision trees to maximize throughput. By prioritizing the avoidance of "high-cost" errors—where picking the wrong radio leads to massive data loss—it achieves up to a 20.83% throughput gain over state-of-the-art systems while being 50% smaller and more interpretable.

What Problem Does This Paper Address?

Existing multi-radio systems like MARS use standard Machine Learning (ML) to decide which radio to use for a transmission. These models typically treat all errors as equal (a "0/1 loss" approach).

However, in radio selection, not all mistakes are created equal. Imagine two scenarios:

  1. Low-cost error: Radio A offers 4600 bps and Radio B offers 4500 bps. If the model picks B, you only lose 100 bps. It’s a mistake, but a tolerable one.
  2. High-cost error: Radio A offers 7000 bps and Radio B offers 2000 bps. If the model picks B here, you lose 5000 bps. This is a disaster for system performance.

The authors found that in previous systems, these "high-cost" errors accounted for a staggering 40% of total throughput loss. Standard ML algorithms struggle here because they usually assign costs to classes (e.g., "always penalize picking Zigbee wrongly"), whereas in IoT, the cost is sample-specific—it depends on the exact signal conditions at that microsecond.

Methods

The researchers developed COMNETS, which stands out through three key technical choices:

  • Cost-Sensitive Tree Alternating Optimization (TAO): Instead of standard decision trees (like CART) that grow greedily and treat all errors the same, COMNETS uses the TAO algorithm. TAO allows the model to incorporate a specific "weight" or "cost" for every single data point in the training set, directly tied to the throughput difference between the radios at that moment.
  • Oblique Decision Trees: Most decision trees are "axis-aligned," meaning they look at one feature at a time (e.g., "Is RSSI > -80?"). COMNETS uses oblique trees, where each node can look at a linear combination of multiple features simultaneously. This makes the model much more powerful and efficient.
  • Sparsity and Stability: By applying $\ell_1$ regularization, they forced the model to use as few features as possible (Sparse Oblique Trees). They also leveraged TAO’s structural stability—unlike standard trees that might change completely if you add a few new data points, TAO-optimized trees remain stable, making them easier for engineers to trust and interpret.

Datasets

The authors didn't just simulate this; they built it. They deployed a 15-node multi-radio mesh network (Raspberry Pi 3Bs with Zigbee and LoRa radios) at two distinct real-world locations:

  • Location A & B: Complex environments with different building materials and human traffic.
  • Traces: They collected a comprehensive dataset of 25,500 data packets, recording features like Hop Number (HN), RSSI (signal strength), and RNP (Required Number of Packets).
  • Ground Truth: Each packet was manually labeled with the "correct" radio choice based on the actual throughput achieved by both radios at that time.

Key Insights and Ideas

  • The "Gray Region" is the Battleground: Radios like Zigbee dominate at short ranges, and LoRa dominates at long ranges. The "gray region" (500-1200m) is where they are competitive, and where smart selection matters most.
  • Interpretability via Sparsity: By looking at the sparse trees, the authors gained a "physical" understanding of the model. They found the model uses Hop Number for "coarse" decisions (how far am I from the gateway?) and RSSI/RNP for "fine-grained" decisions (what is the local interference like right now?).
  • Stability is a Feature: Traditional trees are notoriously "jittery"—small changes in data lead to massive changes in the decision logic. By using TAO, the authors proved that the decision rules remain consistent even as the dataset grows, which is vital for deploying these on resource-constrained IoT hardware.

Key Results

  • Throughput Gains: COMNETS outperformed the previous state-of-the-art (MARS) by 20.83% at Location A and 17.39% at Location B.
  • Efficiency: The decision trees were 50% smaller than standard models, making them much easier to convert into simple IF...ELSE statements for low-power microcontrollers.
  • Latency: In high-traffic scenarios, COMNETS kept latency within the required 55ms bound, whereas previous systems saw latency spikes as they made poor radio choices that led to queuing delays.

Limitations and Future Work

  • Feature Engineering: Currently, the model relies on pre-calculated metrics like RSSI and RNP. The authors suggest that future versions could feed raw packet sequences directly into the model to find patterns that human designers might miss.
  • Hardware Overhead: While the ML model is small, the current hardware (Raspberry Pi) is relatively power-hungry. Optimizing the implementation for even smaller, battery-sipping microcontrollers is a logical next step.
  • Temporal Dynamics: The paper notes that retraining intervals need more study—how often does the model need to "learn" new environmental patterns as seasons change?

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 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 (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 learning for throughput optimization in Multi-radio IoT NETworkS}},
  booktitle = {ETFA},
  year = {2025},
  url = {https://openreview.net/forum?id=HHLbhiRPR0},
}

Loading related content...

Figure