Abstract
Wireless Mesh Networks (WMNs) improve their capacities by equipping mesh nodes with multi-radios tuned to non-overlapping channels. However, data forwarding between nodes faces challenges as link bandwidth varies dynamically. This paper presents a machine learning-based forwarding algorithm that allows nodes to dynamically select the next hop with the highest potential bandwidth. By maintaining only three past status points, nodes can predict future link capacities. Additionally, a geometrical algorithm defines a forwarding region to prevent network flooding. Simulations show that this approach significantly improves transmission speed and outperforms existing algorithms.
Introduction
In Wireless Mesh Networks (WMNs) using cognitive radios, the "shortest path" is rarely the most efficient. Cognitive radios allow nodes to sense their environment and switch channels dynamically to avoid interference. While this increases overall capacity, it introduces significant volatility in link quality. Traditional routing protocols struggle to adapt to these rapid fluctuations.
Researchers have proposed a lightweight, machine-learning-driven solution that predicts future link quality using minimal historical data. By combining these predictions with geometric constraints, the algorithm ensures data moves efficiently toward its destination without overwhelming the network.
Methodology
The proposed system utilizes a two-pronged strategy: Bandwidth Prediction and Geometric Filtering.
Bandwidth Prediction
To minimize computational overhead and battery drain on mobile devices, the algorithm uses a "Rule of Three." Each node tracks only the last three bandwidth changes ($t_0, t_1, t_2$) for its neighboring links. Using polynomial-based prediction, the node calculates a coefficient to estimate the potential bandwidth at a future time. This proactive approach allows nodes to select neighbors with increasing capacity rather than those whose performance is declining.
Geometric Forwarding Region
To avoid the inefficiencies of network flooding, the authors developed a geometric filtering mechanism. Using GPS coordinates, a source node identifies the destination's location and defines a 90-degree "forwarding quadrant"—essentially a cone of progress toward the target.
The node only evaluates neighbors within this quadrant, ensuring that data always moves closer to the destination while maintaining the flexibility to choose the highest-bandwidth link available in that direction.
Evaluation and Results
The algorithm was evaluated in a simulated environment using MATLAB, with node densities ranging from 100 to 300 mobile devices. The "ML Forwarding" approach was compared against Fuzzy Logic and standard Supervised Learning strategies.
Performance Gains
The simulation results demonstrated that the ML-based approach consistently outperformed its peers.
- Reduced Latency: At a density of 300 nodes, the ML Forwarding algorithm achieved a transmission delay of approximately 230ms, compared to 260ms for Supervised Learning and 280ms for Fuzzy Logic.
- Increased Throughput: The transmission speed for the ML approach reached nearly 2900 MB/ms at high node densities, significantly higher than the 1800 MB/ms achieved by Supervised Learning.
Conclusion and Limitations
The proposed algorithm offers a scalable, lightweight solution for routing in volatile wireless environments. By using minimal historical data and geometric constraints, it achieves high throughput with low latency. However, several limitations remain. The current model assumes a "noiseless" environment and relies on consistent GPS availability, which may not be feasible in indoor or high-interference real-world scenarios. Future work will focus on refining the prediction models to handle more chaotic environments while preserving low computational costs.
Summary updated