The paper refers Healthcare Internet of Things (IoT) applications generate time-sensitive data that must be processed close to the point of collection. Cloud-only processing can introduce communication delay, while uneven task placement across fog nodes can create congestion and poor resource utilization. This paper presents a hybrid Whale Optimization Algorithm–Firefly Algorithm (WOA–FA) framework for latency-aware task scheduling in healthcare fog computing. WOA is used to explore the global task-allocation search space and coordinate communication among fog resources; the Firefly Algorithm refines local sensor-side data-collection and task-placement decisions. The workflow initializes candidate solutions, evaluates a latency-oriented objective, updates candidates using the WOA and FA search mechanisms, and retains improved solutions until the stopping condition is reached. The manuscript’s comparison chart reports latency for Throttle, SACO, Round Robin (RR), Proposed, MPSO, and BLA at 500–2500 IoT sensors. At 2500 sensors, the plotted values are approximately 96 for the proposed method, 97 for Throttle, 98 for SACO, 102 for BLA, 105 for MPSO, and 128 for RR. These chart-read values suggest that the proposed approach remains among the lowest-latency methods as the number of sensors increases. Because the supplied manuscript does not provide raw experimental logs, confidence intervals, or a fully specified test environment, the plotted values should be interpreted as approximate figure-derived results rather than independently reproduced measurements.
Physics-Informed neural networks (PINNs) are one of the most active frontiers in scientific machine learning, used to solve the forward and inverse problems governed by partial differential equations. It incorporates governing physical laws and equations into neural network training i.e., loss function. This review discusses twenty-four primary sources, beginning with the biological inspiration and architecture of artificial neural networks, proceeds through the formal construction of PINNs, and then concentrates on the two forces that most strongly govern their training behavior – the placement of collocation points and spectral bias. Different sampling strategies are then reviewed, including residual-based adaptive refinement (RAR), residual-based adaptive refinement with distribution (RAR-D), residual-based adaptive distribution (RAD), residual-based adaptive node generation (RANG), the point-adaptive collocation method PACMANN, residual and gradient adaptive sampling strategies, reinforcement learning-driven sampling, importance sampling, Hessian-based provably accurate sampling, and imbalanced-learning-inspired approaches such as RSmote. We then discuss the neural tangent kernel (NTK), which explains why PINNs sometimes fail to train and why adaptive weighting and self-adaptive mechanisms restore balance among competing loss terms. Finally, we connect this theory to practice through gradient-enhanced PINNs, DeepXDE library and engineering applications in heat transfer and thermal nanoimprint lithography, before identifying open challenges and research directions.
In this paper, we introduce an approach based on Deep Reinforcement Learning to optimize the RIS-NOMA scheme for the 6G wireless network. Some other parameters considered as actions in the proposed algorithm are clustering of users, power distribution, and RIS phase optimization. The two types of deep reinforcement learning algorithms, namely, SAC and MADDPG, were implemented on the proposed architecture. The performance of the proposed framework has been examined based on some parameters including sum rate, SINR, throughput, delay, energy efficiency, and Jain's fairness index. The suggested MADDPG and SAC models have shown better performance than existing models, based on the obtained results.
The fast expansion of IoT systems has led to many problems in ensuring effective and safe data processing with the available computational resources in edge nodes. Conventional techniques used in IoT security systems such as AES need substantial computation power and consume too much energy. In order to overcome this challenge, the presented model applies lightweight cryptography along with the application of the Kalman Filter and adaptive genetic algorithm. The GA-based optimal Kalman filter is able to minimize any possible noise and sudden changes in the sensor data, thus ensuring higher accuracy and reliability of the collected data. According to experimental findings, the proposed model outperforms the classical AES based model. This model saves energy, boosts the speed of processing, increases resilience to attacks, and ensures reliability of data. The suggested MADDPG and SAC models have shown better performance than existing models, based on the obtained results.
Machine unlearning—the selective removal of the influence of specific training data from a deployed machine learning model—has become a pressing problem at the intersection of computational efficiency, regulatory compliance, and AI ethics. Legislative frameworks including the EU’s General Data Protection Regulation (GDPR), the California Consumer Privacy Act (CCPA), and Singapore’s Personal Data Protection Act (PDPA) now require organisations to demonstrate verifiable erasure of personal data from trained AI systems. This review maps the technical landscape of machine unlearning, examining exact methods such as SISA (Sharded, Isolated, Sliced, and Aggregated) training alongside approximate approaches including influence functions, gradient ascent, and model editing. It then situates these methods within the applicable regulatory frameworks and examines the ethical dimensions—fairness, accountability, adversarial misuse, and individual agency—with attention to tensions that remain unresolved. Four open problems emerge clearly: the verification gap, scalability to foundation models, adversarial robustness, and the absence of standardised evaluation benchmarks. Machine unlearning is not merely a compliance mechanism; it is a foundational capability for trustworthy AI, and real progress on it requires genuine convergence across machine learning research, law, and ethics.
