Machine Learning-Based WiFi Indoor Localization with FasterKAN: Optimizing Communication and Signal Accuracy
Abstract
WiFi Received Signal Strength Indicator (RSSI) fingerprint has emerged as a powerful indoor localization technique, surpassing other methods such as GPS-based and Bluetooth-based approaches in terms of accuracy and cost-effectiveness. However, challenges persist in WiFi indoor localization, including the need for shorter latency, simpler systems, better scalability, and robustness. In this study, we applied FasterKAN, a variant of Kolmogorov-Arnold Networks (KAN), to address these challenges. FasterKAN significantly shortens model forward time while enhancing localization accuracy. To evaluate its performance, we applied FasterKAN alongside traditional machine learning algorithms to two benchmark datasets: UJIIndoorLoc and SODIndoorLoc. Our results demonstrate remarkable accuracy: 99% for floor and building classification, and 71% for space ID classification. The mean position error (m) is impressively low at 3.56 for the prediction of coordinates. When using graphics processing unit (GPU) accelerating and central processing unit (CPU) only, the model forward time of FasterKAN was about 8.2 and 3.7 times shorter than that of convolutional neural network (CNN) model respectively, with roughly the same number of total parameters. This research contributes to advancing WiFi indoor localization, providing an accurate, robust, and efficient solution for real-world applications such as navigation in large public spaces, optimizing resource management in smart buildings, enhancing operational efficiency in industrial settings.