Generalization aspect of accurate machine learning models for CSI-based localization.

Area of Science:

• Wireless communication
• Signal processing
• Machine learning

Background:

• Localization is crucial for applications like autonomous driving and IoT.
• Channel State Information (CSI) offers superior data for localization compared to RSSI due to its stability and richness.
• Existing methods require improvement in accuracy and generalization across different environments.

Purpose of the Study:

• To enhance localization accuracy by integrating classical and deep learning techniques using CSI.
• To evaluate the generalization capabilities of Multi-Layer Perceptron Neural Networks (MLP NN) and K-nearest neighbors (KNN) in varied environments.
• To compare the performance of deep learning and classical methods for CSI-based localization.

Main Methods:

• Implemented a Multi-Layer Perceptron Neural Network (MLP NN) for deep learning-based localization.
• Utilized K-nearest neighbors (KNN) as the classical machine learning approach.
• Employed a modified data splitting strategy to rigorously test generalization by minimizing training and test set overlap.

Main Results:

• Both MLP NN and KNN approaches surpassed state-of-the-art performance on the benchmark dataset.
• Localization accuracy for both methods decreased when tested in unseen environments.
• The MLP NN demonstrated superior generalization potential compared to KNN, performing better on data outside the training distribution.

Conclusions:

• Combining classical and deep learning methods with CSI improves localization accuracy.
• Deep learning models, specifically MLP NN, show promise for robust localization systems that generalize well to new environments.
• Further research into deep learning models is warranted to understand and enhance their generalization capabilities in real-world localization challenges.