Articles

Modern cloud computing environments increasingly rely on microservices deployed in containerized clusters to serve highly dynamic workloads. Efficiently managing resources for these microservices is critical to ensure low runtime latency while minimizing memory and other resource usage. Traditional static or rule-based autoscaling methods often over-provision (wasting memory and incurring high cost) or under-provision (causing SLA violations), particularly under rapidly changing demand. In this work, we present a reinforcement learning (RL)-based framework for resource optimization that explicitly leverages high-performance computing (HPC) and parallel/distributed processing capabilities to enable real-time decision-making at large scale. The resource management problem is formulated as a Markov Decision Process (MDP), with a deep RL agent trained to continuously learn optimal scaling and memory allocation policies for hundreds to thousands of containerized microservices. Our architecture distributes RL inference and training across multiple HPC nodes, exploiting multi-core CPUs and GPUs to process large state/action spaces and to evaluate scaling actions with sub-second latency — a requirement unattainable on conventional servers. Experimental results on realistic cloud benchmarks show that our HPC-enabled RL framework achieves up to 25–30% lower average response times and ~15–20% lower memory usage compared to state-of-the-art autoscaling baselines, while maintaining 99% SLA compliance under bursty workloads. These findings demonstrate that coupling RL with HPC infrastructures enables real-time, fine-grained resource optimization in complex microservice ecosystems, aligning directly with the performance and scalability goals of supercomputing research.

The rapid proliferation of artificial intelligence (AI) models has transformed numerous domains; however, their efficient deployment in resource-constrained environments—such as edge and embedded devices— continues to pose substantial challenges. This survey systematically examines contemporary software engineering practices designed to optimize and deploy AI models on hardware with limited computational power, memory, and energy resources. It explores a diverse range of methodologies, including architectural strategies, development toolchains, testing and validation frameworks, edge-tailored MLOps paradigms, and critical security and privacy considerations. By synthesizing insights from recent literature, this paper identifies prevailing challenges, highlights successful approaches, and outlines promising avenues for future research to support robust and scalable AI integration in pervasive low-resource systems. This comprehensive overview aims to serve as a valuable reference for researchers and practitioners navigating the complexities of edge AI development.

The rapid expansion of the Internet of Things has introduced large-scale security and trust challenges due to device heterogeneity, limited resources, and exposure to evolving cyberattacks. Traditional rule-based protection mechanisms and static trust policies often underperform against adaptive attackers and distributed threats. This paper presents a lightweight deep-learning-driven security and trust enhancement framework that combines sequence-based behavioral modeling with an adaptive trust controller to reduce false alarms and enable more reliable device-level decision making. The framework relies on a long short-term memory network for temporal traffic modeling, while continuously updating device trust using anomaly severity, historical consistency, and device criticality. A reproducible preprocessing pipeline is specified to transform heterogeneous network-flow features into fixed-length sequences. The paper also defines an evaluation protocol on widely used intrusion datasets and modern Internet-of- Things-oriented corpora, and reports a structured template for presenting detection performance and edge feasibility metrics, including parameter footprint, inference latency, memory usage, and energy consumption.

Abstract—Retrieval-augmented generation can improve the factual grounding of language models, but generated answers may still contain claims that are only partially supported or unsupported by the retrieved context. Persian has comparatively few resources for developing compact, auditable verification models. This paper presents FaRAG-Guard, a three-way Persian claim verifier that predicts supported, partially supported, or hallucinated and selects an evidence sentence from the source passage. We construct a controlled benchmark from 56,766 answerable PQuAD examples using document-disjoint splits, balanced labels, and template families held out across training, test, and stress conditions. FaRAG-Guard adapts XLM-RoBERTa-base with low-rank adapters, updating 1.04 million of 279.08 million parameters. Across three random seeds, the model obtains 72.06% ± 2.91% macro-F1 and 73.78% ± 1.95% balanced accuracy on the 900-instance main test set, with 72.70% ± 7.95% evidence token-F1. Median single-claim latency is 21.85 ± 3.13 ms on an NVIDIA Tesla T4, with 1.26 GB peak allocated memory. The model exceeds the strongest main-test baseline by 7.58 macro-F1 points; however, its 56.85% ± 15.15% stress-test macro-F1 is below a token-overlap baseline, exposing sensitivity to phrasing shift. The results establish a reproducible compact baseline while identifying robustness and calibration as unresolved requirements.

In modern distributed computing environments, efficient resource allocation remains a challenging problem due to dynamic workloads, heterogeneous resources, and complex inter-task dependencies. Traditional scheduling heuristics (e.g., First-Come-First-Served and Round Robin) often fail to adapt to such conditions or to exploit structural information in job execution graphs. This paper presents an integrated framework for adaptive resource allocation that combines reinforcement learning (RL) with graph neural networks (GNNs) to learn scheduling policies in distributed systems. The GNN component encodes graph-structured system state—such as task dependency graphs and resource availability—into compact representations, enabling the RL agent to generalize across varying workloads and system topologies. Based on these representations, the RL agent selects scheduling actions (e.g., task selection or placement) with the objective of optimizing long-term performance metrics, including throughput, job completion time, resource utilization, and fairness. We describe the design of the proposed RL+GNN framework, covering state and action formulation as well as reward design, and provide a simulation-based conceptual evaluation in a cloud/HPC cluster scheduling setting. The results suggest that incorporating graph-based state representations into reinforcement learning can improve the adaptivity and scalability of resource allocation strategies, particularly in environments with complex dependency structures and dynamic behavior.

Urban transportation systems in megacities like Bangkok face mounting pressure due to population growth, limited infrastructure, and increasing vehicle usage. In this study, we explore the application of modern machine learning (ML) techniques—including Long Short-Term Memory (LSTM) networks, Graph Neural Networks (GNNs), and Transformer-based models—to analyze and predict urban traffic patterns in Bangkok. Using real-world data from traffic sensors, GPS traces, and public transportation records, we construct a data-driven forecasting pipeline that captures complex spatial-temporal dependencies. Our results show that GNN-based models significantly outperform traditional baselines and even LSTM in key metrics, while Transformer models offer complementary advantages. We also discuss the potential for federated learning to support privacy-preserving traffic analytics in Bangkok’s decentralized urban infrastructure. This study contributes a scalable, high-accuracy framework for smart urban mobility planning and real-time traffic prediction in Southeast Asian cities.

This paper applies advanced machine learning methods to predict software defects in the early stages of complex software projects. Statistical data and historical code measurements from the studied project were first collected and preprocessed, and K-means clustering was used to assign initial class labels. Several classification algorithms, including SVM, Random Forest, Naive Bayes and ensemble methods, were then trained on these data. The experiments showed that Random Forest and other hybrid methods performed better, with some models approaching 99% prediction accuracy. Empirical evidence also indicated that machine learning can substantially reduce software testing cost and time while improving final product quality. Challenges remain, including insufficient labeled data and imbalanced datasets. Finally, approaches such as synthetic data, balancing methods and advanced deep models are proposed to enable practical adoption in large industries.

Large Language Models (LLMs) are extremely resource-intensive to deploy, demanding high memory and compute. Static provisioning often leads to waste or unmet demand. We propose a conceptual framework that uses reinforcement learning (RL) and self-adaptive software engineering to optimize resource use in LLM deployments. An RL agent monitors system metrics (throughput, latency, GPU/CPU utilization) and takes actions such as scaling instances, adjusting model precision, or modifying batch sizes. The system employs a Monitor-Analyze-Plan-Execute (MAPE-K) loop where dynamic configuration parameters are tuned online to maximize throughput and minimize cost. We illustrate the approach with examples: RL-driven autoscaling (showing ~40–50% higher GPU utilization) and adaptive inference optimizations like key-value caching (up to 4× speedup). Real- world LLM deployments (cloud services and edge settings) exhibit highly variable workloads; our framework adapts to these changes. Experiments and industry reports show that RL- based adaptation can significantly improve resource efficiency and performance.

Software bug prediction aims to automatically identify code modules likely to contain defects, improving quality and reducing testing costs. Traditional methods use static code metrics (complexity, churn) with classical machine learning, but these often ignore inter-module relationships. Recently, deep learning (DL) techniques have shown promise by learning features from code text and structure. In particular, Graph Neural Networks (GNNs) can model software as graphs (e.g. classes/modules as nodes and dependencies as edges) to capture rich semantics. This paper reviews both traditional and DL-based defect prediction approaches, with emphasis on innovative GNN models using static analysis and code dependency graphs. We summarize that DL models (e.g. LSTM, CNN) often outperform conventional ML, and that GNNs leveraging code structure yield further gains in accuracy and F1 score. In existing studies, LSTM-based classifiers achieved ~87% accuracy, and multi-view GNN approaches report 17–45% relative F1 improvement. These results suggest that combining static metrics with code dependency graphs in GNN frameworks can significantly enhance bug prediction.

As web applications become increasingly complex, comprehensive and effective automated testing is more necessary than ever. This paper reviews and analyzes recent approaches to automated web application testing using artificial intelligence algorithms, including machine learning, deep learning and neural networks. After examining challenges in traditional web testing, such as dynamic forms and complex logic, intelligent methods are introduced in three main groups: (1) machine learning for generating and analyzing test cases; (2) neural networks and deep learning for understanding user interface elements and automated testing; and (3) deep reinforcement learning for autonomous exploration of the web state space. Selected frameworks, including WebRLED and WebExplor, are presented together with their experimental results. The conclusion indicates that AI-based methods can substantially improve code coverage and fault detection rates. Future research directions include integrating large language models (LLMs) and expanding access to real-world datasets for modern training.

Software systems today face increasing demands for high performance and efficient memory usage. Traditional optimization methods, often hard-coded and inflexible, struggle to adapt to complex and dynamic workloads. Machine Learning (ML) algorithms offer a promising approach by automatically learning patterns and making intelligent decisions to optimize execution time (runtime) and memory consumption. This paper provides a comprehensive review of recent research (2021–2025) on applying ML techniques – including neural networks, reinforcement learning, and evolutionary algorithms – to performance optimization in software systems. We discuss how ML-driven solutions have achieved significant improvements, such as reducing program execution time, enhancing memory/cache efficiency, and intelligently allocating resources in cloud environments.

Complex problems typically exhibit nonlinear, multi-criteria and big-data characteristics, making a single artificial intelligence technique insufficient to solve them. This paper examines the importance of hybrid AI methods for improving the accuracy and efficiency of complex problem solving. A hybrid approach combines several AI methods simultaneously or sequentially within an integrated framework. Various studies have shown that human–machine collaboration, or hybrid intelligence, improves machine learning model accuracy, while metaheuristic algorithms in hybrid frameworks can accelerate neural network convergence. Empirical findings also indicate that hybrid AI models significantly improve prediction accuracy for complex problems (Akbari et al., 1400 in the Iranian calendar).

The Internet of Things (IoT) is a network of smart devices that exchange data extensively. Its rapid growth brings significant security and privacy challenges alongside many benefits. Traditional cybersecurity methods are insufficient for protecting the IoT ecosystem because of centralized structures and vulnerability to emerging threats. Blockchain, with decentralization, transparency, immutability and distributed consensus, has emerged as a new approach to improving IoT security and trust. Using a literature-based and analytical research method, this paper examines blockchain’s role in improving security and privacy in multilayer IoT architectures. After reviewing security challenges, blockchain capabilities for addressing them are explained. A conceptual framework integrating the two technologies is presented, and recent research findings are analyzed. The results indicate that blockchain can make data tampering harder, strengthen identity management and restrict unauthorized access. Future directions include lightweight consensus algorithms and improved scalability.

The rapid scaling of large language models (LLMs) has led to unprecedented computational costs and environmental impact. We address the problem of multi-objective optimization for LLM training, balancing model performance against resource usage (training time, energy, GPU-hours, carbon footprint). We survey and evaluate state-of- the-art techniques—model pruning, quantization, knowledge distillation, neural architecture search (NAS), and hyperparameter tuning via evolutionary or reinforcement learning—in terms of their trade-offs between accuracy (or loss) and efficiency. Using recent experimental data from public benchmarks (e.g. BERT fine-tuning on GLUE tasks, GPT-family training), we analyze how each method shapes the Pareto frontier of accuracy vs. cost. For example, static 8-bit quantization has been shown to cut energy use by ~29% with negligible accuracy loss, while structured pruning can speed up inference by ~63% for minor accuracy degradation. Advanced LLM compression methods achieve even larger gains: a multi-objective shift-add reparameterization method achieved over 80% reduction in memory and energy usage compared to full models. We include Pareto-plots (accuracy vs energy) to visualize these trade-offs (Figure 1–4). Overall, we find that multi-objective search (e.g. Bayesian optimization or genetic algorithms) can systematically identify configurations that lie on the Pareto-optimal front, enabling practitioners to choose models that best fit their constraints. Our paper highlights that, consistent with “Green AI” principles, moderate sacrifices in accuracy can yield large efficiency gains, and we provide actionable recommendations for training sustainable LLMs.

Edge computing brings computation closer to data sources, reducing latency but posing new energy challenges for battery-powered and distributed resources. In this work, we present a machine learning (ML)-based framework that dynamically optimizes energy use in heterogeneous edge computing environments. Our approach incorporates adaptive workload distribution, hardware-aware ML model design, and low-power inference techniques. The framework uses learning algorithms (e.g. reinforcement learning) to allocate tasks across edge nodes and cloud resources based on current load and energy profiles. We evaluate the framework in a representative IoT scenario, showing that ML-guided scheduling can reduce energy consumption by ~30–50% compared to static allocation strategies while meeting latency constraints. Key innovations include multi-objective optimization of computational and hardware parameters (inspired by DynaSplit) and the use of model compression and accelerators to enable low-power inference. These results demonstrate the promise of ML- driven resource management for sustainable edge computing.

With recent advances in large language models (LLMs), automatic code completion has become one of their key applications in software development. This paper examines automatic code evolution with a focus on Python. Advanced language models such as GPT-4 and OpenAI Codex, and tools such as GitHub Copilot, are reviewed, and their code completion and automatic code generation capabilities are analyzed. Research findings indicate that LLMs have substantially improved code completion accuracy and speed and, in some cases, developer productivity; for example, a 55% reduction in task completion time has been reported (Perry et al., 2023). However, challenges such as insecure code generation and the need for human oversight remain (Pearce et al., 2022). Alongside a review of domestic and international research, the paper describes methods for evaluating these models and presents findings on their impact on Python programming. It concludes by summarizing opportunities and challenges in adopting LLM-based programming assistants.