Autoencoder Extreme Learning Machine for Fingerprint-Based Positioning: A Good Weight Initialization is Decisive

Abstract

Indoor positioning based on machine-learning (ML) models has attracted widespread interest in the last few years, given its high performance and usability. Supervised, semisupervised, and unsupervised models have thus been widely used in this field, not only to estimate the user position, but also to compress, clean, and denoise fingerprinting datasets. Some scholars have focused on developing, improving, and optimizing ML models to provide accurate solutions to the end user. This article introduces a novel method to initialize the input weights in autoencoder extreme learning machine (AE-ELM), namely factorized input data (FID), which is based on the normalized form of the orthogonal component of the input data. AE-ELM with FID weight initialization is used to efficiently reduce the radio map. Once the dimensionality of the dataset is reduced, we use k-nearest neighbors to perform the position estimation. This research work includes a comparative analysis with several traditional ways to initialize the input weights in AE-ELM, showing that FID provide a significantly better reconstruction error. Finally, we perform an assessment with 13 indoor positioning datasets collected from different buildings and in different countries. We show that the dimensionality of the datasets can be reduced more than 11 times on average, while the positioning error suffers only a small increment of 15% (on average) in comparison to the baseline.