Enforcing the forgetting (or unlearning) of specific portions of a neural network's (NN) memory is necessary for both security and ethical considerations. Moreover, targeted forgetting can enhance the network's performance by eliminating the learning derived from compromised data. Conventional approaches, such as retraining the model from scratch, are associated with significant time and energy costs, while alternative methods can degrade the neural network's performance in terms of complexity and predictive accuracy. To address these challenges, this work introduces two novel approaches that mitigate the drawbacks associated with traditional unlearning methods. Both approaches leverage the Long Short-Term Memory (LSTM) network, chosen for its superior learning quality and efficient processing time—crucial factors in such algorithms. The first approach involves modifying the LSTM architecture by coarse tune the weight matrix of the forget gate to facilitate controlled forgetting. The second approach follows the idea of intruders by inducing poluuted data to the system, and the idea is based on learning the model through corrupted/bad data, which causes the LSTM memory to be changed at specific time intervals that correspond to the data that should be forgotten. We present a performance evaluation of the second approach, with a future study planned for the performance assessment of the first approach. The evaluation demonstrates the superiority of the proposed solutions in terms of quality of service and implementation simplicity.

Offline Reinforcement Learning (RL) methods constrain the policy to align with the behaviour policy, mitigating extrapolation errors caused by out-of-distribution (OOD) actions. Implicit Q-Learning (IQL), a popular offline RL algorithm, leverages expectile regression and introduces an in-sample learning paradigm that enhances the policy evaluation stage without querying OOD actions. However, the crucial parameter$T$for expectile regression in IQL is fixed, limiting both its performance and flexibility across diverse datasets. In this paper, we propose Cos-IQL, an improved IQL approach based on cosine similarity, which optimizes the policy evaluation function by measuring the cosine similarity between the policy and the behaviour policy. Cos-IQL is essentially a multi-step offline RL algorithm but retains the advantages of in-sample learning, thus avoiding the risks of OOD actions. In addition, Cos-IQL can adaptively adjust parameter$T$without elaborate fine-tuning. We evaluate Cos-IQL on D4RL benchmark datasets and compare its performance against recent competitive offline RL algorithms. Experimental results show that Cos-IQL achieves state-of-the-art performance.

Balancing sustainable energy with cellular deployment can be challenging. However, multi-user MIMO antennas for dynamic positioning can benefit both objectives. This study presents an optimization algorithm leveraging dynamic drone deployment to alleviate cellular communication congestion in densely populated areas. Two deployment strategies are examined: prioritizing energy efficiency by minimizing antenna footprint, and optimizing service quality by placing serving offloading drones. A novel Voronoi planning tool models both approaches, aiding in estimating energy consumption across various scenarios, particularly in the context of modern multi-user massive MIMO and beamforming techniques. A comparative energy analysis for different techniques is conducted, along with establishing a theoretical minimal coverage antenna footprint.

Interest in applying Reinforcement Learning (RL) to Autonomous Vehicles (AVs) is experiencing a rapid and substantial expansion. Proximal Policy Optimization (PPO), a well-known RL algorithm with two versions, is simple to implement and has a high level of generality. In this paper, we first analyze the issues in each of the original PPO versions: asymmetric penalty in the Adaptive KL Penalty Coefficient PPO version, gradient loss and pessimistic estimate in the Clipped PPO version. Therefore, we propose three improved PPO algorithms: Adaptive JS Penalty Coefficient PPO, Leaky PPO, and Parametric PPO. To validate the effectiveness of the proposed algorithm, we generated three autonomous driving scenarios in the Metadrive simulator. Experimental results demonstrate that Leaky PPO outperforms the other five PPO variant algorithms in various autonomous driving simulation scenarios. Furthermore, we demonstrate that the Leaky PPO outperforms other popular RL algorithms and achieves state-of-the-art performance.

The evolution of vehicular networks continues with the advent of LPWAN (Low Power Wide Area Network). Thus, several studies currently concern the integration of LPWANs with ITS (Intelligent Transportation Systems). Among these LPWANs, LoRa with LoRaWAN, with its coverage capabilities, support of mobility, and implementation cost, is in the spotlight. However, the latter presents performance problems in highly mobile environments. Therefore, optimizing mobility in LoRaWAN is imperative for its integration into ITS. This requires performance improvements such as loss containment in very dense environments. In this paper, a resource reservation approach is proposed. It is based on device trajectory prediction and traffic differentiation in a dense multi-operator environment. This mechanism aims at reserving resources before the arrival of the device on the predicted antenna, thus reducing the rejections at the join phase while favoring “its subscribers”.

Two models are proposed to optimize unmanned aerial vehicles (UAVs) traffic offloading in cellular networks. The first model optimizes the realtime traffic service, while delaying non-realtime traffic in the cell buffers (BS and the currently serving drone). Delayed traffic is then transmitted later when free resources are available and has a maximum service delay limit. The proposed model provides a heuristic solution to minimize the losses in non-realtime traffic based on the maximum delay and the size of the cell buffers. The second model completes the work by providing an optimal Integer Linear Programming solution to minimize the number of needed UAV assuming no data loss and buffering availability. It is also parameterized with the same delay limits as the first model. The performance of the proposed models is studied using the call detail record (CDR) dataset of the city of Milan cellular network provided by Telecom Italia. The two models showed great QoS performance based on the maximum delay of non-realtime traffic and cell buffer size compared to models that do not include the buffering and the delay limits.

Base station (BS) capacity for cellular networks is fixed and therefore cannot be easily adapted to support temporary high demands due to unusual traffic (demonstration, sales, competitions, etc.). Unmanned Ariel Vehicle (UAV)/drone can be used as an additional temporary bandwidth (BW) provider capable of covering multiple cells.Two placement approaches are investigated to optimize UAV deployment for this purpose. The first adopts a rule based approach without cooperation: we place UAVs in the cells where the exceeding demand is the highest, allowing each UAV to only serve a single cell at a time. In the second algorithm, relying on the Knapsack problem, we design a solution where UAVs cooperate by sharing a portion of their available BW with neighbouring cells in order to cover all the demand with the lowest possible amount of drones. Their performance is compared using two months of real dataset from the Milan Cellular Network provided by Telecom Italia.

The development of vehicular technologies and infrastructures leads to development in mobility handling for wireless communications. Improving connectivity establishment and reliability became an issue, especially for vehicles that may move out of antenna coverage during connection establishment. This paper focuses on improving LoRaWAN connectivity for roaming devices by combining a machine learning predictor and DNS prefetching to gather information necessary for connection establishment before the device comes under coverage, thus reducing the overall latency for connection establishment. The paper also relates to other issues by comparing the solution with other approaches and studying antenna occupation.

IoT technologies currently operate as independent silos, and roaming is possible only if there are prior interconnection agreements. To our knowledge, there are no standardised procedures for interconnecting different IoT networks for roaming. The focus of IoTRoam is to set up an operational roaming model that scales, seamlessly works with existing IoT infrastructures and interconnects on a global basis with minimum initial configuration requirements. As a Proof-of-Concept, we designed, implemented and tested a roaming LoRaWAN architecture using time-tested infrastructures on the Internet such as PKI and the DNS. The IoTRoam experience helped us to propose changes to the LoRaWAN Backend Interface Specification that have since been accepted. We also evaluated whether the proposed mechanisms satisfy constrained IoT requirements.

LPWANs are networks characterized by the scarcity of their radio resources and their limited payload size. To extend the efficiency of the data transmission by decreasing the traffic sent from sensors, this paper proposes a lossy compression method using known ML techniques. We embedded a pre-trained neural network directly on constrained LoRaWAN devices and we tested the trade-off between compression ratio and accuracy of the compression algorithm. This paper studies multiple aspects of the system - energy consumption, error rate due to the lossy compression, compression ratio and the impact of LSTM parameter quantization - to measure the possible strengths and weaknesses of using a dual prediction system in order to reduce transmission costs. Surprisingly, machine learning used in this context does not consume a lot of energy and it even leads to energy saving in the very constrained devices which are the sensors.

LPWANs are networks characterised by the scarcity of their radio resources and their limited payload size. LoRaWAN offers an open, easy-to-deploy and efficient solution to operate a long-range network. To efficiently communicate using IPv6, the LPWAN working group from the IETF developed a solution called Static Context Header Compression (SCHC). It uses context rules, which are linked to a given End Device, to compress the IPv6 and UDP header. Since there may be a huge variety of End Devices profile, it makes sense to store the rules remotely and use a system to retrieve the profiles dynamically. In this paper we propose a mechanism based on DNS to find the context rules associated with an End Device, allowing it to be downloaded from an HTTP Server. We evaluate the corresponding delay added to the communications using experimental measurements from a real testbed.

We propose a solution for Electric Vehicles (EVs) energy management in smart cities, where a deep learning approach is used to enhance the energy consumption of electric vehicles by trajectory and delay predictions. Two Recurrent Neural Networks are adapted and trained on 60 days of urban traffic. The trained networks show precise prediction of trajectory and delay, even for long prediction intervals. An algorithm is designed and applied on well known energy models for traction and air conditioning. We show how it can prevent from a battery exhaustion. Experimental results combining both RNN and energy models demonstrate the efficiency of the proposed solution in terms of route trajectory and delay prediction, enhancing the energy management.

Virtual Content Delivery Network (vCDN) orchestration is necessary to optimize the use of resources and improve the performance of the overall SDN/NFV-based CDN function in terms of network operator cost reduction and high streaming quality. It requires intelligent and enticed joint SDN/NFV orchestration algorithm due to the evident huge amount of traffic to be delivered to end customers of the network. In this paper, a global vCDN architecture and an exact approach for finding the optimal path orchestration(s) and vCDN component instantiation(s) (OCPA) are proposed. Moreover, several scenarios are considered to quantify the OCPA behavior and to compare its efficiency in terms of caching and streaming cost, orchestration time, vCDN replication number, and other cost factors. Then, it is implemented and evaluated under different deployment flavors. Several scenarios are considered to study the algorithm's behavior and to quantify the impact of both network and system parameters.

Real Call Detail Records (CDR) are analyzed and classified based on Support Vector Machine (SVM) algorithm. The daily classification results in three traffic classes. We use two different algorithms, K-means and SVM to check the classification efficiency. A second support vector regression (SVR) based algorithm is built to make an online prediction of traffic load using the history of CDRs. Then, these algorithms will be integrated to a network planning tool which will help cellular operators on planning optimally their access network.

In a smart grid environment, we study the coalition formation of prosumers that aim at entering the energy market. It is paramount for the grid operation that the energy producers are able to sustain the grid demand in terms of stability and a minimum production requirement. We design an algorithm that seeks to form coalitions that will meet both of these requirements: a minimum energy level for the coalitions and a steady production level, which leads to finding uncorrelated sources of energy to form a coalition. We propose an algorithm that uses graph tools such as correlation graphs or clique percolation to form coalitions that meet such complex constraints. We validate the algorithm against a random procedure and show that, it not only performs better in terms of social welfare for the power grid, but also that it is more robust against unforeseen production variations due to changing weather conditions for instance.

For several years, works on cognitive radio have been published allowing to reuse TV Whitespaces. On another hand, many new results appear on user and crowd mobility due to big data recent developments, especially from cellular traces. We propose here a method for cellular resource planning taking into account user mobility. Since users move, the bandwidth resource should also move accordingly. We design a score based method using TV Whitespace, and user experience, to determine from which cell it should be removed and to which one it should be added. Combined with traffic history it calculates scores for each cell. Bandwidth is reallocated on a half day basis. Before that, real traces of cellular networks in urban districts are presented which confirm that static network planning is no longer optimal. A dynamic femtocell architecture is then presented. It is based on mesh interconnected elements and designed to serve the score based bandwidth allocation algorithm. The score method along with the architecture are simulated and results are presented. They confirm the expected improvement in bandwidth and delay per user while maintaining a low operation cost at the operator side.

In this paper, we propose an autonomous self-organization mechanism for the monitoring of the cold chain. All along the transportation through this logistic chain, the sensors are moved with the goods in very different networks. We argue that depending on the situation some protocols are more suited than others and it is necessary the sensors to adapt and to switch dynamically depending on where they are. The mechanism we designed allows a sensor to autonomously detect the context where it is and to choose the best protocols for this situation. It assures the adaptability to different contexts where a sensor may move (warehouses, trucks,...). It implements a hybrid synchronization mechanism to assure duty cycle synchronization between nodes with high power efficiency and context adaptation. This hybrid mechanism uses wake up signals to activate a sleeping sensor to assure fast synchronization and efficient power consumption optimization. This mechanism also changes dynamically the routing protocol to the different networks the sensors enter in (dense or sparse topologies, with or without total visibility among all nodes,...).

In this paper, we propose a Hybrid Multi-Channel Multi-hop Medium Access Control (HMM-MAC) meeting the safety and non-safety requirements in Vehicular Ad-hoc Networks when no infrastructure is present (V2V communication) with only a single transceiver at each vehicle. The simulation results showed that our scheme could achieve a higher delivery ratio of road safety messages and a high probability to select a free transmission channel in a certain region and during a certain time.

In WSN, it is not often desirable to use the GPS technology. Indeed, the use of GPS is expensive and may reduce the overall network performance. Moreover, indoor reception of the GPS signal is not possible. The Link Quality Indicator (LQI) is defined in the 802.15.4 standard, but its context of use is not specified in this standard. Some works on the LQI, few of which are field experiments, have shown that the LQI decreases as the distance increases. However, the challenge of clustering mechanisms is to form the smallest number of clusters by maximizing distances separating cluster heads to provide an efficient cover of the network and also minimizes the cluster overlaps. This reduces the amount of channel contention between clusters, and also improves the efficiency of algorithms that run at the level of the cluster heads. Therefore we propose an analytical model based on the use of the LQI in order to derive an optimally one-dominating set in which the smaller distance separating two cluster heads is improved.

It is proved that the normal suite of the TCP/IP protocols works well when it is used in terrestrial networks transporting real time applications. The situation is quite different when wireless networks are used for the same services; their susceptibility to errors, losses as well as the existing delay cause a degradation in the overall performance. Due to the existing demand to interconnect terrestrial networks with satellite communications using DVB-S2/DVB-RCS standards, the satellite link poses a challenge which is to overcome the problems mentioned when applications as voice over IP (VOIP) are considered. In the same context, cross-layer mechanisms are being used successfully in ad-hoc networks. They propose interactions between layers different from the traditional ones in the protocol stack. These mechanisms provide a fast and better way to exchange parameters, drawing as a result a good option to improve the performance in wireless networks. This work proposes an innovative algorithm to adapt dynamically the coding for VOIP applications over satellite using a cross layer interaction between the application and network layers in order to improve the throughput and the perceived quality. This interaction with the network layer indirectly takes into account the state of the MAC and physical layers.