Section: New Results

Protocols and Models for Wireless Networks - Application to VANETs

Connection-less IoT - Protocol and models

Participants : Iman Hemdoush, Cédric Adjih, Paul Mühlethaler.

The goal is to construct some next-generation access protocols, for the IoT (or alternately for vehicular networks). One starting point are methods from the family of Non-Orthogonal Multiple Access (NOMA), where multiple transmissions can "collide" but can still be recovered - with sophisticated multiple access protocols (MAC) that take the physical layer/channel into account. One such example is the family of the Coded Slotted Aloha methods.Another direction is represented by some vehicular communications where vehicles communicate directly with each other without necessarily going through the infrastructure. This is also true more generally in any wireless network where the control is relaxed (such as in unlicensed IoT networks like LoRa). One observation is that in such distributed scenarios, explicit or implicit forms of signaling (with sensing, messaging, etc.), can be used for designing sophisticated protocols - including using machine learning techniques.

During this study, some of the following tools should be used: protocol/algorithm design (ensuring properties by construction), simulations (ns-2, ns-3, matlab, ...) on detailed or simplified network models, mathematical modeling (stochastic geometry, etc...) ; machine-learning techniques or modeling as code-on-graphs.

The first result we have obtained concerns Irregular Repetition Slotted Aloha (IRSA) which is a modern method of random access for packet networks that is based on repeating transmitted packets, and on successive interference cancellation at the receiver. In classical idealized settings of slotted random access protocols (where slotted ALOHA achieves $1/e$), it has been shown that IRSA could asymptotically achieve the maximal throughput of 1 packet per slot. Additionally, IRSA had previously been studied for many different variants and settings, including the case where the receiver is equipped with “multiple-packet reception” (MPR) capability. We extensively revisit the case of IRSA with MPR. We present a method to compute optimal IRSA degree distributions with a given maximum degree n. A tighter bound for the load threshold ($G/K$) was proven, showing that plain K-IRSA cannot reach the asymptotic known bound $G/K=1$ for $K>1$, and we prove a new, lower bound for its performance. Numerical results illustrate that optimal degree distributions can approach this bound. Second, we analyze the error floor behavior of K-IRSA and provide an insightful approximation of the packet loss rate at low loads, and show its excellent performance. Third, we show how to formulate the search for the appropriate parameters of IRSA as an optimization problem, and how to solve it efficiently. By doing that for a comprehensive set of parameters, and by providing this work with simulations, we give numerical results that shed light on the performance of IRSA with MPR. A final open question is: what is the impact of introducing more structure in the slot selection (like Spatially Coupled Coded Slotted Aloha) and how best to do so?

Indoor positionning using Channel State Information (CSI) from a MIMO antenna

Participants : Abdallah Sobehy, Paul Muhlethaler, Eric Renault ( Telecom Sud-Paris ).

The channel status information is used for locating a node by applying machine learning [35] techniques. We propose a novel lightweight deep learning solution to the indoor positioning problem based on noise and dimensionality reduction of MIMO Channel State Information (CSI): real and imaginary parts of the signal received. Based on preliminary data analysis, the magnitude of the CSI is selected as the input feature for a Multilayer Perceptron (MLP) neural network. Polynomial regression is then applied to batches of data points to filter noise and reduce input dimensionality by a factor of 14. The MLP’s hyper-parameters are empirically tuned to achieve the highest accuracy. The method is applied to a CSI dataset estimated at an 8 x 2 MIMO antenna that is published by the organizers of the Communication Theory Workshop Indoor Positioning Competition. The proposed solution is compared with a state-of-the-art method presented by the authors who designed the MIMO antenna that is used to generate the data-set. Our method yields a mean error which is 8 times less than that of its counterpart. We conclude that the arithmetic mean and standard deviation misrepresent the results since the errors follow a log- normal distribution. The mean of the log error distribution of our method translates to a mean error as low as 1.5 cm. We have shown that, using a K-nearest neighbor learning method an even better, indoor positioning is achieved. The input feature is the magnitude component of CSI which is pre-processed to reduce noise and allow for a quicker search. The Euclidean distance between CSI is the criterion chosen for measuring the closeness between samples. The proposed method is compared with three other methods, all based on deep learning approaches and tested with the same data-set. The K-nearest neighbor method presented in this paper achieves a Mean Square Error (MSE) of 2.4 cm, which outperforms its counterparts.

Predicting Vehicles Positions using Roadside Units: a Machine-Learning Approach

Participants : Samia Bouzefrane ( Cnam ), Soumya Banerjee ( Birla Institute Of Technology, Mesra ), Paul Mühlethaler, Mamoudou Sangare.

We study positioning systems using Vehicular Ad Hoc Networks (VANETs) to predict the position of vehicles. We use the reception power of the packets received by the Road Side Units (RSUs) and sent by the vehicles on the roads. In fact, the reception power is strongly influenced by the distance between a vehicle and a RSU. We have already used and compared three widely recognized techniques : K Nearest Neighbors (KNN), Support Vector Machine (SVM) and Random Forest. We have studied these techniques in various configurations and discuss their respective advantages and drawbacks. We revisit the positioning problem VANETs but we also consider Neural Networks (NN) to predict the position [22]. The neural scheme we have tested in this paper consists of one hidden layer with three neurons. To boost this technique we use an ensemble neural network with 50 elements built with a bagging algorithm. The numerical experiments presented in this contribution confirm that a precise prediction can only be obtained when there is a main direct path of propagation. The prediction is altered when the training is incomplete or less precise but the precision remains acceptable. In contrast, with Rayleigh fading, the accuracy obtained is much less striking. We observe that the Neural Network is nearly always the best approach. With a direct path the ranking is: Neural Network, Random Forest, KNN and SVM except in the case when we have no measurement in [30m; 105m] where the ranking is Neural Network, Random Forest, SVM and KNN. When there is no direct path, the ranking is SVM, NN, RF and KNN but the difference in performance between SVM and NN is small.

Combining random access TDMA scheduling strategies for vehicular ad hoc networks

Participants : Fouzi Boukhalfa, Mohamed Hadded ( Vedecom ), Paul Mühlethaler, Oyunchimeg Shagdar ( Vedecom ).

This work is based on Fouzi Boukhalfa's PhD which started in October 2018, [29],[15]. The idea is to combine TDMA protocols with random access techniques to benefit from the advantages of both techniques. Fouzi Boukhalfa proposes to combine the DTMAC protocol introduced by Mohamed Hadded with a generalization of CSMA. This generalized CSMA uses active signaling; the idea is to send signaling bursts in order to select a unique transmitter. The protocol that Fouzi Boukhalfa obtains reduces the access and merging collisions of DTMAC but can also propose access with low latency for emergency traffic. The idea is that vehicles access their slots reserved with DTMAC but the transmission slots encompass a special section at the beginning with active signaling. The transmission of the signaling burst, during a mini-slot, is organized according to a random binary key. A '1' in the key means that a signaling burst will be transmitted, while a '0' means that the vehicle senses the channel on this mini-slot to potentially find the transmission of a signaling burst by another vehicle. Fouzi Boukhalfa shows that if we use a random key to transmit the signaling burst it very significantly decreases the collision rate (both merging and access collisions) and that emergency traffic can have a very small access delay. Fouzi Boukhalfa builds an analytical model which thoroughly confirms the simulation result. This model can encompass detection error in the selection process of the signaling bursts. It is shown that with a reasonable error rate the performance is only marginaly affected.

Forecasting traffic accidents in VANETs

Participants : Samia Bouzefrane ( Cnam ), Soumya Banerjee ( Birla Institute Of Technology, Mesra ), Paul Mühlethaler, Mamoudou Sangare.

Road traffic accidents have become a major cause of death. With increasing urbanization and populations, the volume of vehicles has increased exponentially. As a result, traffic accident forecasting and the identification of the accident prone areas can help reduce the risk of traffic accidents and improve the overall life expectancy.

Conventional traffic forecasting techniques use either a Gaussian Mixture Model (GMM) or a Support Vector Classifier (SVC) to model accident features. A GMM on the one hand requires large amount of data and is computationally inexpensive, SVC on the other hand performs well with less data but is computationally expensive. We present a prediction model that combines the two approaches for the purpose of forecasting traffic accidents. A hybrid approach is proposed, which incorporates the advantages of both the generative (GMM) and the discriminant model (SVC). Raw feature samples are divided into three categories: those representing accidents with no injuries, accidents with non incapacitating injuries and those with incapacitating injuries. The output or the accident severity class was divided into three major categories namely: no injury in the accident, non-incapacitating injury in the accident and an incapacitating injury in the accident. A hybrid classifier is proposed which combines the descriptive strength of the baseline Gaussian mixture model (GMM) with the high performance classification capabilities of the support vector classifier (SVC). A new approach is introduced using the mean vectors obtained from the GMM model as input to the SVC. The model was supported with data pre-processing and re-sampling to convert the data points into suitable form and avoid any kind of biasing in the results. Feature importance ranking was also performed to choose relevant attributes with respect to accident severity. This hybrid model successfully takes advantage of both models and obtained a better accuracy than the baseline GMM model. The radial basis kernel outperforms the linear kernel by achieving an accuracy of 85.53%. Data analytics performed including the area under the receiver operating characteristics curve (AUC-ROC) and area under the precision/recall curve(AUC-PR) indicate the successful application of this model in traffic accident forecasting. Experimental results show that the proposed model can significantly improve the performance of accident prediction. Improvements of up to 24% are reported in the accuracy as compared to the baseline statistical model (GMM). The data about circumstances of personal injury in road accidents, the types of vehicles involved and the consequential casualties were obtained from data.govt.uk.

Although a significant improvement in accuracy has been observed, this study has several limitations. The first concerns the dataset used. This research is based on a road traffic accident dataset from the year of 2017 which contains very few data samples for the no injury and non-incapacitating injury types of accident. The data was unbalanced not just with respect to the output class but also with respect to the sub features of various attributes. Moreover, aggregating the accident severity into just three categories limits the scope of the study and the results obtained. The greater the number of severity classes, the less is the amount of extra training data required to feed in the SVC to avoid overfitting. Thus, datasets with sufficient records corresponding to each class are desirable and must be used for further study.

The second limitation concerns the dependence of the SVC model on parameters and attribute selection. In this study, the performance of SVC relies heavily on the feature selection results and the mean vectors obtained from the GMM. In order to improve the accuracy of the support vector classifier, other approaches like particle swarm optimization (PSO), ant colony optimization, genetic algorithms etc. could be used for effective parameter selection. In addition to this, more kernels like the polynomial kernel and the sigmoid kernel could be tested to improve future model performances.