Evaluating Link Lifetime Prediction to Support Computational Offloading Decision in VANETs

Paulo Rocha, Alisson Barbosa de Souza, Gilvan Maia, César L. C. Mattos, Francisco Airton Silva, Paulo A. L. Rego, Tuan Anh Nguyen, Jae-Woo Lee2022 · DOI

TL;DR

This paper tackles the challenge of accurately predicting link lifetimes in Vehicular Ad Hoc Networks (VANETs) to improve computational offloading decisions. By employing and evaluating various machine learning regression models, the authors demonstrate a significant reduction in prediction errors compared to traditional methods, leading to more efficient offloading strategies.

What Problem Does This Paper Address?

Vehicular Ad Hoc Networks (VANETs) enable a range of intelligent transportation systems, from collision alerts to autonomous driving. Many of these applications require significant computational power, leading to the concept of "computational offloading," where tasks are sent to nearby vehicles or edge servers. However, VANETs are characterized by high vehicle mobility, resulting in short and dynamic communication link lifetimes. This makes it difficult to accurately predict when a link will break, which is crucial for making timely and effective offloading decisions. Existing methods for link lifetime prediction often suffer from low accuracy, limited feature sets, and a narrow focus on specific scenarios (e.g., highways only), hindering their practical application in diverse urban and high-traffic environments.

Methods

The core of this research lies in evaluating and applying machine learning (ML) regression algorithms for link lifetime prediction. The authors explored several ML models, including Support Vector Regression (SVR), XGBoost, MLP, KNN, and Random Forest. They developed a methodology that involves:

  • Scenario Generation: Creating realistic urban and highway simulation scenarios using SUMO and NS-3.
  • Data Collection: Gathering contextual traffic data from these simulations.
  • Feature Engineering: Identifying and extracting eight key features that influence link lifetime, including distance, speed, acceleration, pseudo-angle, and Spatial Locality Similarity (SLS).
  • Data Cleaning and Preprocessing: Normalizing features and handling noisy data.
  • Model Training and Evaluation: Training the selected ML models and evaluating their performance using metrics like Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE).
  • Integration with Offloading: Modifying an existing computational offloading algorithm (GCF) by replacing its link lifetime prediction function with the best-performing ML model (initially SVR, later XGBoost).

A significant contribution is the creation of new datasets tailored to different traffic densities (low, medium, high) within urban scenarios, allowing for more context-aware model training.

Datasets

The paper utilizes datasets generated from simulations of two primary scenarios:

  • Urban Scenario: A 2 km section of downtown Manhattan, New York, USA.
  • Highway Scenario: A 5 km section of the BR-116 highway in Fortaleza, Brazil.

Additionally, the authors introduce new datasets specifically for urban environments, categorized by traffic density (low, medium, high). These datasets are made publicly available.

Key Insights and Ideas

  • ML Outperforms Traditional Methods: Machine learning regression models significantly reduce prediction errors for link lifetime compared to conventional methods like LLT, leading to more accurate estimations.
  • Feature Importance: The inclusion of a broader set of features, particularly contextual ones like pseudo-angle and SLS, is crucial for improving prediction accuracy.
  • Context-Aware Prediction is Key: Training ML models on datasets specific to traffic density and location (urban vs. highway, or even within urban zones) yields better performance than generic models.
  • SVR and XGBoost Shine: SVR and XGBoost emerged as strong contenders, offering a good balance between prediction accuracy and computational overhead.
  • Offloading Efficiency Gains: By improving link lifetime prediction, computational offloading strategies can achieve higher success rates and lower recovery rates, enhancing overall application performance and quality of service.

Key Results

  • Reduced Prediction Error: ML models, particularly SVR and XGBoost, achieved significantly lower MAE, MAPE, and RMSE values compared to the traditional LLT method in both highway and urban scenarios. For instance, in the highway scenario, MLP achieved an MAE of 1.89s, a substantial improvement over LLT's 24.14s.
  • Faster Decision Times: While some ML models (like KNN and Random Forest) introduced significant overhead, SVR demonstrated a favorable decision time, making it suitable for real-time offloading decisions.
  • Improved Offloading Performance: The GCFML approach (GCF with ML-based prediction) showed improved offloading success rates and reduced recovery rates compared to the original GCF and other baseline algorithms, especially in high-density urban and highway scenarios. For example, in the highway scenario, GCFML achieved higher success rates and lower recovery rates across various densities.
  • Density-Specific Models: In urban scenarios, models trained on density-specific datasets (GCFML Density and GCFML Zone) showed better performance than a merged dataset, with the GCFML Zone (context-aware based on local density) demonstrating the most promising results in reducing recovery rates.

Limitations and Future Work

  • Simulation-Based: The study relies on simulations, and real-world validation is needed to confirm the findings.
  • Specific Scenarios: While urban and highway scenarios were considered, other environments might present unique challenges.
  • Computational Overhead: Some ML models still incur significant computational overhead, which might be a concern for resource-constrained vehicles.
  • Future Work: The authors suggest exploring new prediction hypotheses, testing models trained on more precise situations, and investigating whether scenario-specific models can outperform more generic ones. They also aim to further optimize the offloading process by considering more complex scenarios and real-world constraints.

Summary updated

BibTeX
@misc{rocha2022evaluating,
  author = {Paulo Rocha and Alisson Barbosa de Souza and Gilvan Maia and César L. C. Mattos and Francisco Airton Silva and Paulo A. L. Rego and Tuan Anh Nguyen and Jae-Woo Lee},
  title = {{Evaluating Link Lifetime Prediction to Support Computational Offloading Decision in VANETs}},
  year = {2022},
  doi = {10.3390/s22166038},
}
Figure