A Data-driven Transfer Learning Method for Radio Map Estimation

2018 IEEE Wireless Communications and Networking Conference (WCNC), 2018

Fifth generation wireless networks (5G) will face key challenges caused by diverse patterns of traffic demands and massive deployment of heterogeneous access points. In order to handle this complexity, machine learning techniques are expected to play a major role. However, due to the large space of parameters related to network optimization, collecting data to train models for all possible network configurations can be prohibitive. In this paper, we analyze the possibility of performing a knowledge transfer, in which a machine learning model trained on a particular network configuration is used to predict a quantity of interest in a new, unknown setting. We focus on the tilt-dependent received signal strength maps as quantities of interest and we analyze two cases where the knowledge acquired for a particular antenna tilt setting is transferred to (i) a different tilt configuration of the same antenna or (ii) a different antenna with the same tilt configuration. Promising results supporting knowledge transfer are obtained through extensive experiments conducted using different machine learning models on a real dataset.

2021

With outstanding features, Machine Learning (ML) has become the backbone of numerous applications in wireless networks. However, the conventional ML approaches face many challenges in practical implementation, such as the lack of labeled data, the constantly changing wireless environments, the long training process, and the limited capacity of wireless devices. These challenges, if not addressed, can impede the effectiveness and applicability of ML in wireless networks. To address these problems, Transfer Learning (TL) has recently emerged to be a promising solution. The core idea of TL is to leverage and synthesize distilled knowledge from similar tasks as well as from valuable experiences accumulated from the past to facilitate the learning of new problems. Doing so, TL techniques can reduce the dependence on labeled data, improve the learning speed, and enhance the ML methods' robustness to different wireless environments. This article aims to provide a comprehensive survey on applications of TL in wireless networks. Particularly, we first provide an overview of TL including formal definitions, classification, and various types of TL techniques. We then discuss diverse TL approaches proposed to address emerging issues in wireless networks. The issues include spectrum management, localization, signal recognition, security, human activity recognition and caching, which are all important to nextgeneration networks such as 5G and beyond. Finally, we highlight important challenges, open issues, and future research directions of TL in future wireless networks.

2021 16th Annual Conference on Wireless On-demand Network Systems and Services Conference (WONS), 2021

The use of artificial intelligence is foreseen to be pervasive in future mobile radio networks, enabling dynamic and proactive radio resource provisioning and allocation as well as end-to-end optimization of the network architecture. Current approaches in mobile radio networks commonly assume having a complete batch of data on the specific network element when optimizing and adapting the network working configuration. Such a pipeline is at odds with the increasing complexity and extreme flexibility of 5G and next generation systems where reconfiguration decisions might be taken rather frequently, and with only few data available. In this paper, we focus on the problem of predicting channel quality and average number of active user equipment when a limited amount of data is available from the cell to predict and a high number of predictions need to be carried out simultaneously. We propose a transfer learning framework based on one dimensional convolutional neural networks and explore several models with different complexity overhead for the prediction task across 100 cells. The performance of the proposed framework is validated against classical machine learning approaches in terms of accuracy and computation time when varying the amount of data available for training. Achieved results indicate that transfer learning outperforms the "non-transfer" approaches, specifically when the amount of data available from the cell to predict is scarce. Index Terms-Channel quality prediction, active users prediction, deep transfer learning, time series forecasting *At the time of writing this paper Claudia Parera was with Politecnico di Milano and Nokia Bell Labs.

The WiFi-based indoor localization problem (WILP) aims to detect the location of a client device given the signals received from various access points. WILP is a complex and very important task for many AI and ubiquitous computing applications. A major approach to solving this task is through machine learning, where upto-date labeled training data are required in a large scale indoor environment. In this paper, we identify WILP as a transfer learning problem, because the WiFi data are highly dependent on contextual changes. We show that WILP can be modeled as a transfer learning problem for regression modeling, where we identify several important cases of knowledge transfer that range from transferring the localization models over time, across space and across client devices. We also share our working experience in WILP and transfer learning research in a realistic problem solving setting, and discuss a data set we have made public for advancing this research.

IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 2000

We present a new machine learning framework for indoor location estimation. In many cases, locations could be easily estimated using various traditional positioning methods and conventional machine learning approaches based on signalling devices, e.g., access points (APs). When there exist environmental changes, however, such traditional methods cannot be employed due to data distribution change. In order to circumvent this difficulty, we introduce feature relevance network-based method, which focuses on interrelatedness among features. Feature relevance networks are connected graphs representing concurrency of the signalling devices such as APs. In the newly created relevance network, a test instance and the prototype of a location are expanded until convergence. The expansion cost corresponds to distance between the test instance and the prototype. Unlike other methods, our model is nonparametric making no assumptions about signal distributions. The proposed method is applied to the 2007 IEEE International Conference on Data Mining Data Mining Contest Task #2 (transfer learning), which is a typical example situation where the training and test datasets have been gathered during different periods. Using the proposed method, we accomplish the estimation accuracy of 0.3238, which is better than the best result of the contest.

international conference on telecommunications, 2021

This paper continues previous explorations in the area of deep learning applications in the field of cellular wireless networks, specifically the problem of identifying optimal beams in a highly directional urban environment, using topographical data. In our previous work, we have studied the problem and demonstrated how deep-learning can be used on static topographical data for prediction of optimal beams. In this paper, we show a potential architecture for realization of the same for a network of nodes in a given area, taking into account challenges of computational complexity, response time and the inherent architecture of the next generation RAN. This is achieved by using deep transfer learning as a way of translating between a global feature space inherent to the coverage area and local variations thereof, specific to the location of each radio-unit.

Internet Technology Letters, 2020

Indoor localization algorithms in smart cities often use Wi-Fi fingerprints as a database of Received Signal Strength (RSS) and its corresponding position coordinate for position estimation. However, the issue of fingerprinting is the use of different platform-devices. To this end, we propose a Long Short-Term Memory (LSTM)-based novel indoor positioning mechanism in smart city environment. We used LSTM, a type of recurrent neural network to process sequential data of users' trajectory in indoor buildings. The proposed approach first utilizes a database of normalizing fingerprint landmarks to calculate WiFi Access Points (WAPs) RSS values to mitigate the fluctuation issue and then apply the normalization parameters on the RSS values during the online phase. Afterwards, we constructed a transfer model to adapt the RSS values during the offline phase and then applying it on the RSS values from the different smartphones during the online phase. Thorough simulation results confirm that the proposed approach can obtain 1.5 to 2 meters positioning accuracy for indoor environments, which is 60 % higher than traditional approaches.

IEEE Access

The performance of any Mobile Wireless Network (MWN) is dependent on the appropriate level of radio coverage, with Path Loss (PL) models being a valuable resource for its evaluation. Recently, advancements in Machine Learning (ML) and Deep Neural Networks (DNNs) have been applied to radio propagation to produce new data-driven PL models. Notoriously, these advancements have also allowed the inclusion of non-classical inputs, such as satellite images. However, data-driven PL models are often developed under the assumption that training and test data distributions are similar, which is a weak assumption in real-world scenarios. Thus, generalization (i.e., the model's ability to perform on different data distributions) is a crucial aspect of data-driven PL models in the context of Mobile Network Operators (MNOs). This paper proposes a new data-driven PL model, the Ubiquitous Satellite Aided Radio Propagation (USARP) model, developed to enhance the geographical generalization capabilities of empirical PL models, by using satellite images. The USARP model considers self-supervised learning to extract general data representations of the radio environment from satellite images, improving the PL prediction Root Mean Square Error (RMSE) of the 3 rd Generation Partnership Project (3GPP) PL model in the order of 9 dB, and for a data distribution distinct from the training data. Moreover, it was demonstrated the potential of the USARP model in terms of geographical and radio environment generalization. Although the generalization capabilities of ML regression algorithms are limited, the chosen USARP architecture and the use of regularization techniques had a positive impact on its geographical generalization performance. INDEX TERMS Wireless networks, radio propagation, path loss models, satellite data, deep learning, selfsupervised learning, convolutional neural networks.

2021 17th International Conference on Network and Service Management (CNSM), 2021

Recently, transfer learning (TL) has emerged as a powerful machine learning method in distributed environments. Transferring the knowledge between distributed agents helps reduce both learning time and computing costs. However, in a communication system, the advantage of TL comes with communication costs. To make an optimal decision of transfer between two agents, we try to answer three key questions: i) which information should be transferred from a source to a target?, ii) how this transferred information will be adapted to the target? and iii) when should TL be triggered to optimize the costs?. To this end, we introduce a new concept of similarity based on the Best Approximation Theory and a general transfer rule. Then, we propose a model to evaluate the feasibility and optimality of TL. We verify our proposed model in the context of the wireless channel selection problem using contextual multiarmed bandits. Experimental results show optimal TL decisions can be made, and Extra Action is an efficient technique for TL in channel selection.

IEEE Access

Accurate channel model predictions are crucial in mobile communication systems to identify the coverage area of cellular base stations. It also allows network operators to optimally choose the locations of the new sites, solve coverage gap problems and optimize the current network parameters. Current prediction models use ray tracing techniques that are too computationally expensive and depend on the 3D maps, which are costly and need to be regularly updated. This paper proposes a multi-modal channel model prediction algorithm using satellite images to extract the environmental features and other numerical features. For an accurate evaluation, experimental measurements in the 2100 MHz band are gathered and combined with 2D maps from two different LTE network areas with varying characteristics to practically reference our results and compare the results with the ray-tracing output. Using the well-known AlexNet architecture as a baseline for our model and introducing new numerical features, we achieve a mean absolute error (MAE) of 2.06 dB and 2.6 root-mean-square error (RMSE) with 4.8 dB enhancement compared to only using numerical features. Using transfer learning, we train the model in area one and test it in another area. We achieve 1.47 dB MAE and 1.99 RMSE with 77.34% reduction in the training time. INDEX TERMS Channel model prediction, coverage estimation, LTE network, deep learning.