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

Abstract

Localization is the process of determining the position of an entity in a given coordinate system. Due to its wide range of applications (e.g. autonomous driving, Internet-of-Things), it has gained much focus from the industry and academia. Channel State Information (CSI) has overtaken Received Signal Strength Indicator (RSSI) to achieve localization given its temporal stability and rich information. In this paper, we extend our previous work by combining classical and deep learning methods in an attempt to improve the localization accuracy using CSI. We then test the generalization aspect of both approaches in different environments by splitting the training and test sets such that their intersection is reduced when compared with uniform random splitting. The deep learning approach is a Multi Layer Perceptron Neural Network (MLP NN) and the classical machine learning method is based on K-nearest neighbors (KNN). The estimation results of both approaches outperform state-of-the-art performance on the same dataset. We illustrate that while the accuracy of both approaches deteriorates when tested for generalization, deep learning exhibits a higher potential to perform better beyond the training set. This conclusion supports recent state-of-the-art attempts to understand the behaviour of deep learning models.

Supervised Contrastive CSI Representation Learning for Massive MIMO Positioning

“…As compared with the true positions shown in Fig. 5 a), we achieve a mean positioning error of 35.2 m, which is better than the lowest errors reported in [6], [24] and [21] as 42 m, 40 m and 37 m respectively. In [6], its best result was achieved by converting the Nadaraya-Watson estimator to a three-layer neural network and using a SVD-based similarity metric.…”

“…To train the encoder f θ pHq and evaluate its positioning capability, we randomly split the data into a training and a test set. The ratio ρ of test samples is set to be 10% (with 498 samples) or 20% (with 996 samples), similar to the settings in [21] and [6].…”

“…Other handcrafted CSI similarity metrics have been designed for general MIMO systems. For example, average Euclidean distance between CSI magnitude values over multiple antennas and sub-carriers is used in [4], [21]. Furthermore, the left singular vector corresponding to the largest singular value of MIMO channel frequency response matrix is selected as the CSI feature in [6], and inner product of such features is used to measure the CSI similarity; Such a similarity metric is closely related to Chordal distance [22] on the Grassmannian manifold.…”

“…Direct mapping via DNN tries to learn a function p " ProjpCSIq to predict the position p from CSI, which inherently describes the uncertain environment and RF characteristics. However, to train reliable DNNs, dense CSI samples are needed [18]- [20], and their positioning accuracies are inferior to classical kNN methods as indicated in [21], [23], [24], especially when the training CSI samples are sparse.…”

Application of Reinforcement Learning and Deep Learning in Multiple-Input and Multiple-Output (MIMO) Systems

“…Accuracy of localization using CSI is improved by combing multi-layer perceptron NN and K-nearest neighbors techniques in [ 177 ]. Both schemes then tested for generalization aspect in different scenarios by dividing the training and validation data in a sense that the intersection is minimized as compared to the uniform random splitting.…”