Bishmita Hazarika

Unmanned aerial vehicles (UAVs) are essential for providing communication and computation services in disaster recovery scenarios where traditional infrastructure is compromised. However, challenges related to energy efficiency, real-time adaptability, coverage, load balancing, and safe navigation persist, particularly in dynamic disaster environments. In this study, we propose a comprehensive framework that integrates Generative AI (GenAI) with graph neural networks (GNN) to dynamically generate hover points for waypoint-based UAV navigation and realistic task generation based on environmental conditions. The GNN-based collision avoidance mechanism further ensures safe navigation by allowing UAVs to avoid obstacles and no-fly zones while coordinating with neighboring UAVs in real-time. To optimize UAV swarm operations, we introduce a multi-agent graph reinforcement learning (MAGRL) framework, enabling UAVs to maximize overall system utility by refining hover point selection, task allocation, and load balancing in response to environmental changes. A graph attention mechanism enhances UAV coordination, improving communication efficiency and decision-making. Extensive simulations show that the proposed GenAI-GNN and MAGRL framework significantly outperforms existing methods in task completion, energy efficiency, and overall system utility in disaster recovery scenarios.

Achieving real-time decision-making and efficient resource management in dynamic, large-scale Internet-of-Vehicles (IoV) networks is a significant challenge due to their inherent complexity and scale. To address this, we propose a digital twin (DT)-assisted IoV framework that integrates a novel semi-synchronous adaptive federated learning (AdFL) approach with multiagent deep reinforcement learning, enhanced by generative artificial intelligence (GenAI) techniques, specifically conditional variational autoencoders (CVAEs). The framework optimizes partial task offloading across distributed mobile-edge computing (MEC) servers, ensuring scalable, efficient, and accurate decision-making in heterogeneous vehicular networks. By continuously mirroring the real-time states of vehicles and roadside units (RSUs), the DT framework enables precise resource allocation and adaptive task management. To tackle the complexities of dynamic environments, we design a global model with transformer layers embedded in the federated learning (FL) process, capturing long-range dependencies. A novel semi-synchronous aggregation mechanism is introduced to balance timely updates with model quality. The proposed adaptive federated multiagent reinforcement learning (AF-MARL) algorithm facilitates decentralized, collaborative learning among vehicles and RSUs, optimizing overall cost, and energy efficiency, reducing delay, and improving task completion rates. Extensive simulations demonstrate the effectiveness of the proposed framework against other existing approaches, highlighting its potential to transform real-time decision-making in IoV networks.

In the rapidly evolving domain of vehicular metaverse, this study introduces a cutting-edge quantum-based decentralized and heterogeneity-aware federated learning framework for vehicular metaverse named QV-FEDCOM, which stands as a testament to the innovative fusion of quantum computing principles with federated learning (FL). This framework is ingeniously tailored to address the challenges in a vehicular metaverse, offering a cost-efficient and adaptive solution for the dynamic vehicular landscape. QV-FEDCOM is strengthened by key components like quantum sequential-training-program, with reinforcement learning-based dynamic mode switching to reduce communication costs and manage vehicle states adaptively, and the quantum vehicle-context-grouping utilizing hierarchical clustering and simulated annealing for effective vehicle grouping based on contextual data similarity, addressing the complexities of data heterogeneity. Additionally, the integration of quantum-inspired principal component analysis (Q-PCA) enhances memory efficiency, further optimizing the framework. These elements converge in the QV-FEDCOM algorithm, establishing a decentralized, efficient, and context-aware quantum federated learning (QFL) process that redefines learning dynamics in the vehicular metaverse. Our study also introduces an innovative quantum trajectory loss (QTL) function, specifically designed for trajectory prediction tasks, which combines the Huber loss with an angular deviation penalty to robustly handle errors and penalize large deviations in the predicted trajectory angle. The effectiveness of the QV-FEDCOM framework is rigorously validated through comprehensive simulations, with its performance meticulously compared against various adaptations, showcasing its transformative capabilities within the vehicular metaverse ecosystem.

Vehicular crowdsensing (VCS) within the social Internet of Vehicles (IoV) significantly advances urban transportation management by enhancing road safety, traffic efficiency, and the overall driving experience. This paper presents an intelligent multi-agent deep reinforcement learning (DRL) framework for augmented dynamic task prioritization in a multi-incentive VCS system. Our framework, named intelligent multi-agent reinforcement learning (IMARL), leverages augmented intelligence to integrate human-like decision-making processes with autonomous vehicle operations, ensuring more adaptive and robust task management. The proposed IMARL framework offers several key advantages: it dynamically adjusts the sensing levels of each vehicle, ensuring efficient energy usage and minimized processing times, and employs a data-quality aware multi-incentive utility model to capture both functional and social incentives. Additionally, our framework incorporates a layered server architecture, enhancing system resilience and scalability. Simulation results demonstrate the superiority of our approach. IMARL achieves significant improvements in task completion rates, energy consumption, and processing delays compared to other DRL and non-DRL benchmark methods. Furthermore, our approach exhibits strong adaptability to changing environmental conditions, maintaining high performance even in high-density traffic scenarios. These quantified results validate the effectiveness of the proposed framework, highlighting its potential to significantly enhance VCS systems in real-world applications.

In this study, we present a hybrid deep reinforcement learning (DRL) algorithm, trained using vehicular federated learning (VFL), specifically tailored for dynamic vehicular networks with historical data. Our approach utilizes VFL-based DRL to refine the caching scheme in these networks, focusing on predicting and storing the most effective content nearby to enhance cache efficiency and reduce content request delays. We propose a modified proximal policy optimization (mPPO)-based approach for the DRL-based decision making for caching management, which combines the advantages of proximal policy optimization (PPO) and double deep Q-network (DDQN). Our study encompasses a vehicular framework that includes a central edge node (CEN), roadside units (RSUs), unmanned aerial vehicles (UAVs), and vehicles equipped with historical data. We tackle the challenges posed by varying vehicle density and mobility, nonuniform RSU coverage, and constrained caching capacity. Through comprehensive simulations, we demonstrate that the mPPO outperforms the conventional DRL methods like PPO and DDQN, as well as heuristic approaches. These results underscore the efficacy of the VFL-based mPPO in dynamic vehicular networks, confirming its potential as a viable solution for real-world applications.

With 5G technology evolving, Open Radio Access Network (O-RAN) solutions are becoming crucial, especially for handling the diverse Quality of Service (QoS) needs in vehicular networks. These networks are dynamic and have many different applications, calling for effective O-RAN strategies. This paper examines a three-tier hierarchical O-RAN slicing model, created to address the unique challenges of vehicular networks. The top-level follow 3GPP standards like ultra-reliable and low-latency communications (URLLC), enhanced mobile broadband (eMBB), and massive machine-type communications (mMTC). The middle level is organized by vehicle types, and the lowest level is designed for specific vehicle applications. This approach leads to better network resource management. Additionally, this study explores the advantages of a federated deep reinforcement learning (DRL) approach for efficient learning while maintaining privacy. It introduces a federated DRL approach incorporating federated averaging and deep deterministic policy gradient (DDPG) techniques, to enhance inter-slice operations and resource allocation in vehicular networks. Lastly, the effectiveness of this algorithm is demonstrated through a small simulation in a vehicular framework.

In this study, we investigate the performance of a hybrid reconfigurable intelligent surface (RIS)-assisted non-orthogonal multiple access (NOMA) network, augmented with backscattering capabilities, designed to facilitate integrated sensing and communication (ISAC). Our primary objective is two-fold: first, to enhance the overall communication throughput, and second, to strengthen the sensing power for target detection. To achieve these goals, we introduce two novel dynamic user clustering algorithms namely composite distance and angle-based (CDA) and channel-oriented adaptive (COA) algorithm for grouping users into clusters with fixed base station and RIS positions, where the successive interference cancellation (SIC) is employed for effective communication within each pair. Moreover, we present a comprehensive optimization problem that jointly maximizes the sum rate and sensing power. This problem involves optimizing the transmit beamformer at the base station, the power allocation factors within each cluster, and the phase shifts at the RIS. This methodology not only adheres to strict power constraints and quality of service requirements at each receiving node but also ensures equitable resource allocation among the targets and enforces unit modulus phase shifts at each RIS element. To tackle the complex interdependencies and non-convex nature of the optimization problem, we introduce an advanced iterative algorithm based on alternative optimization (AO). This state-of-the-art technique employs successive convex approximation (SCA) to systematically address this multifaceted problem. Finally, the simulation results empirically validate the proposed algorithm’s effectiveness, considering the number of RIS elements, maximum power budget, number of targets, and imperfect channel state information (CSI) while showing the trade-off between communication and sensing performance.

In this study, we present a novel approach for efficient resource allocation in a digital twin (DT) framework for task offloading in a UAV-aided Internet-of-Vehicles (IoV) network. Our approach incorporates a hybrid machine learning approach that combines asynchronous federated learning (AFL) and multi-agent deep reinforcement learning (DRL) to jointly optimize task completion rate, energy consumption, and delay parameters, enhancing overall system efficiency. We instantiate a DT infrastructure within a UAV-assisted IoV network for V2V and V2I task offloading with three task processing modes and three types of tasks. The DT network is composed of three distinct DTs: task vehicles, service vehicles, and roadside units. Subsequently, we formulate an optimization problem aimed at maximizing the system efficiency while concurrently minimizing delay and total energy consumption. To address this challenging non-convex problem, we introduce a multi-agent DRL algorithm named MARS for resource allocation within the DT-assisted IoV network. This innovative algorithm, MARS, is trained to utilize a hybrid AFL approach referred to as HAFL. MARS optimizes the allocation of resources across various modes of computation, striving to maximize the system's overall utility. Finally, our proposed approach's effectiveness is validated through comprehensive simulation results, where it is compared against various benchmark schemes for evaluation.

This article considers an internet of vehicles (IoV) network, where multi-access edge computing (MAEC) servers are deployed at base stations (BSs) aided by multiple reconfigurable intelligent surfaces (RISs) for both uplink and downlink transmission. An intelligent task offloading methodology is designed to optimize the resource allocation scheme in the vehicular network which is based on the state of criticality of the network and the priority and size of tasks. We then develop a multi-agent deep reinforcement learning (MA-DRL) framework using the Markov game for optimizing the task offloading decision strategy. The proposed algorithm maximizes the mean utility of the IoV network and improves communication quality. Extensive numerical results were performed that demonstrate that the RIS-assisted IoV network using the proposed MA-DRL algorithm achieves higher utility than current state-of-the art networks (not aided by RISs) and other baseline DRL algorithms, namely soft actor-critic (SAC), deep deterministic policy gradient (DDPG), twin delayed DDPG (TD3). The proposed method improves the offloading data rate of the tasks, reduces the mean delay and ensures that a higher percentage of offloaded tasks are completed compared to that of other DRL-based and non-RIS-assisted IoV frameworks.

Digital twin (DT) has emerged as a promising technology for improving resource allocation decisions in Internet of Vehicles (IoV) networks. In this paper, we consider an IoV network where mobile edge computing (MEC) servers are deployed at the roadside units (RSUs). The IoV network provides ubiquitous connections even in areas uncovered by RSUs with the assistance of unmanned aerial vehicles (UAVs) which can act as a relay between RSUs and task vehicles. A virtual representation of the IoV network is established in the aerial network as DT which captures the dynamics of the entities of the physical network in real-time in order to perform efficient resource allocation for delay-intolerant tasks. We investigate an intelligent delay-sensitive task offloading scheme for the dynamic vehicular environment which provides computation resources via local execution, vehicle-to-vehicle (V2V), and vehicle-to-roadside-unit (V2I) offloading modes based on the energy consumption of the system. Moreover, we also propose a multi-network deep reinforcement learning (DRL)-based resource allocation algorithm (RADiT) in the DT-assisted network for maximizing the utility of the IoV network while optimizing the task offloading strategy. Further, we compare the performance of the proposed algorithm with and without the presence of V2V computation mode. RADiT is further evaluated by comparing it with another benchmark DRL algorithm called soft actor-critic (SAC) and a non-DRL approach called greedy. Finally, simulations are performed to demonstrate that the utility of the proposed RADiT algorithm is higher under every condition compared to its respective conditions in SAC and greedy approach. Consequently, the proposed framework jointly improves energy efficiency and reduces the overall delay of the network. The proposed algorithm with UAV relay further increases the efficiency of the network by increasing the task completion rate.