Search Results (24)

Large language models (LLMs) incur prohibitive computational costs when deployed as monolithic systems for multi-domain query processing. This paper proposes Accelerated LLM, a modular architecture that replaces a single general-purpose LLM with an ensemble of task-specialised small language models (SLMs) governed by a neural query router and a Mamdani fuzzy inference system. The router embeds each user query using a frozen sentence encoder and classifies it across four task domains—summarisation, translation, question answering, and text generation—routing confident queries directly to the corresponding SLM. Ambiguous queries are escalated to a three-input fuzzy logic system operating on Query Length, inter-Domain Overlap Score, and Classifier Confidence, enabling principled handling of imprecise inputs. A reinforcement-learning feedback loop, validated through a controlled pilot deployment, continuously refines the routing policy. The complete pipeline, including the sentence encoder, totals approximately 2.14 billion parameters—a 98.8% reduction relative to GPT-3.5 (175 B). The integration of fuzzy logic into the routing stage raises classification accuracy from 91.5% to 94.3% and reduces the hallucination rate to 9.8% (minor) and 6.4% (major). Evaluated on healthcare-augmented benchmarks against ChatGPT-3.5, Claude, Mistral 70B, and two contemporary compact models (GPT-4o-mini and Llama 3.1-8B-Instruct), Accelerated LLM achieves competitive or superior task-specific performance at a fraction of the parameter count. A small-scale pilot evaluation in the legal domain indicates that the routing and fuzzy logic components retain partial effectiveness beyond the primary healthcare setting, though full multi-domain validation remains future work. Full article

Low Earth orbit (LEO) mega-constellations demand scalable routing that survives time-varying topologies, constrained onboard resources, and dynamic traffic. Deterministic shortest-path routing guarantees optimal paths but adapts poorly to real-time loads; distributed deep reinforcement learning (DRL) can introduce loops and inconsistent end-to-end decisions. This paper fuses deterministic inter-domain planning with DRL-based intra-domain forwarding in a single hierarchical framework. An evolutionary greedy algorithm partitions the constellation into compact domains. Dijkstra’s algorithm then computes backbone paths on the domain-level graph. Inside each domain, a context-enhanced Per-Action Dueling Double Deep Q-Network encodes individual neighbors through a weight-shared encoder and summarizes the valid-neighbor set via masked mean pooling. This design lets the policy compare a candidate against the current alternative set without injecting input-order bias. Local one- and two-hop topological features drive decentralized inference. A greedy–beam–Dijkstra fallback ladder guarantees reachability whenever the subgraph stays connected. On a 1584-satellite Starlink Gen1-1 topology, all 21 domain sizes and six inter-domain strategies reach 100% of test pairs; the best average hop count sits at 1.16× the global Dijkstra benchmark. Under an identical 52-dimensional state and training pipeline on 1000 held-out source–destination pairs, Context Per-Action uses 75.8% fewer parameters than a flat multilayer perceptron (MLP), lifts greedy success from 74.6% to 83.5%, and lifts greedy-plus-beam success from 88.3% to 94.5% (means over three independent training seeds). Centralized load-aware routing under dynamic traffic cuts high-load packet loss from 34–73% to 0–9.5% in the adopted flow-level model and preserves 99.2% reachability despite 30% link failures. Zero-shot transfer from ideal Walker topologies to real two-line element (TLE) snapshots and purely local load adaptation remain open; multi-snapshot training or online adaptation is the necessary next step. Full article

To mitigate high expert-annotation costs, domain-preference misalignment, and the inherent trade-off between sensitive-data protection and training utility in petrochemical dataset construction, an iterative framework combining human-feedback-aligned reinforcement learning (RLHF) with post hoc data de-identification is proposed. Direct scoring and pairwise preference feedback are generated using two high-capability language models. A reward model is subsequently trained via a joint Bradley–Terry and mean-squared-error loss, followed by three rounds of closed-loop proximal policy optimization (PPO) constrained by a fixed supervised fine-tuning reference model. Retained high-quality samples are then processed through a four-stage post-RLHF de-identification pipeline. Experimental results demonstrate that the PPO-V3 model achieves a reward score increase of 1.183 over the baseline alongside a 96.2% pairwise win rate, while the sensitivity-aware adaptive differential privacy with context-aware token-level injection (SA-ADP-CTI) post-RLHF de-identification method attains a composite score of 0.9636. The reliability of both the automated feedback and privacy-preservation mechanisms is further validated through blind expert review and manual spot checks. Full article

Robotic piano playing is a challenging benchmark for biomimetic dexterous manipulation, requiring precise timing, coordinated multi-finger motion, and stable key contact. This study proposes a robotic piano-playing framework based on a two-stage reinforcement learning curriculum. Musical Instrument Digital Interface (MIDI) data are converted into target-key and fingering grids to provide future musical goals for policy learning in a parallel MJLab simulation environment. A Soft Actor–Critic (SAC) agent takes a 2106-dimensional observation vector, including joint states, previous actions, musical phase, future key targets, fingering assignments, and piano-key states, and outputs a 21-dimensional continuous action vector for wrist, finger, and global hand-positioning control. Stage 1 weakens physical regularization to facilitate key-pressing acquisition, whereas Stage 2 strengthens power, velocity, acceleration, collision, posture, and finger-speed constraints to improve the regularity of policy outputs and readiness for real-world deployment. Simulation experiments on 30 s right-hand excerpts from Für Elise, Canon, and Beethoven’s Symphony No. 5 achieve frame-wise key-state F1 scores above 0.99 on the first two excerpts and approximately 0.945 on Beethoven. Real-world deployment uses open-loop playback of policy-generated high-level trajectories with low-level joint-position feedback and achieves F1 scores of 0.95, 0.91, and 0.83, respectively, while reproducing representative piano techniques such as chords, octaves, mixed black-and-white-key patterns, overlapping finger actions, and rapid sequential movements. These physical results demonstrate the feasibility of the proposed sim-to-real pipeline for complete 30 s executions; they are not intended as a statistical repeatability study. The results further show that biomimetic robotic hands can learn complex piano-playing skills from MIDI-based task objectives without relying on human motion demonstration trajectories. Full article

A reinforcement learning hyper-heuristic framework for multi-objective scientific workflow scheduling selects among five meta-heuristic optimisers and tunes their control parameters under a Nash Social Welfare reward over makespan, cost, security, and resource utilisation. In its base form it treats security as a static virtual-machine attribute and admits all candidates unconditionally. This paper contributes the mechanism design required to integrate two security-realism layers into the scheduling loop without redesigning the reward: a five-stage zero-trust admission pipeline, a bounded non-stationary per-machine dynamic trust signal, a state-vector augmentation that exposes trust to the agent, and a coupling that attenuates the effective security level seen by the security utility. The two layers act at distinct timescales: admission is a provisioning-time gate on a machine’s structural compliance, whereas the trust signal evolves per decision epoch for the machines already admitted, so static admission and dynamic trust coexist by construction. We evaluate three hyper-heuristic agents on 20 Pegasus workflow instances under both a trust suite and a trust-free baseline. The base framework establishes a sharp separation between the hyper-heuristic and direct task-to-machine RL families; we treat this as an inherited property and ask a different question: can the two security-realism layers be integrated without disturbing it? Across 20 Pegasus instances and three HH-RL agents, the family-level separation is preserved. Under a reproducible evaluation protocol—five independently seeded repeats of the full paired comparison, 100 greedy inference episodes per (agent, workflow, suite) cell, with per-workflow deltas averaged across repeats before testing—the trust extension imposes a small, heterogeneous absorption cost: the median per-workflow shift in Nash reward is −0.24, −0.24, and −0.05 for PDQN, DQNHH, and QLHH respectively, an order of magnitude below the absolute reward levels. The shift is statistically significant for DQNHH (two-sided Wilcoxon p = 0.0014, rank-biserial r = −0.77), marginal for PDQN (p = 0.058), and absent for QLHH (p = 0.18). The security utility stays above 0.91 on every instance, and the family-level scaling robustness is preserved intact. The contribution is therefore a drop-in mechanism whose cost is bounded and small relative to the between-family separation—with a robust workflow-level heterogeneity: the parameterised agent converts the trust signal into consistent gains on the largest DAGs (mean +1.28 on Sipht_1000 and Inspiral_1000 across the five repeats) while paying a small cost on typical instances. Full article

Traditional MBSE faces challenges in verifying the dynamic performance of autonomous systems due to a semantic gap between static architecture and executable algorithms. To address this limitation, this study proposes a system-of-systems (SoS) modeling methodology that connects the Unified Architecture Framework (UAF) with multi-agent reinforcement learning. A Drivers–Challenges–Opportunities–Goals (DCOG) framework maps strategic intent to operational capabilities, which are formalized in the UAF. Driven by the UAF Domain Metamodel (DMM), a pipeline transforms behavioral and resource specifications into a Markov Decision Process (MDP). An Automatic Curriculum Learning Multi-Agent Proximal Policy Optimization (AutoCL-MAPPO) algorithm then resolves the MDP. The autonomous fleet achieves a 73% mission success rate, outperforming the standard MAPPO baseline (65%). These performance metrics are averaged over 1000 independent test episodes to ensure statistical significance, with baseline algorithms evaluated under identical environmental conditions. Using Systems Modeling Language (SysML) as a verification carrier, activity simulations confirm that the generated decision sequences conform to UAF structural logic. Metric constraint deviation analysis provides empirical feedback for iterative design refinement. This methodology establishes a verifiable digital thread, closing the loop between architecture modeling and learned behavior for autonomous SoS and Human–AI Teaming. Full article

In construction project management, the execution phase relies on deterministic scheduling methods that cannot represent activity-level uncertainty or support quantitative corrective-action testing before physical commitment. This study presents a conceptual framework and proof-of-concept application that addresses this gap by integrating Building Information Modeling (BIM) with the COSMOS Simulator—a construction-specific discrete-event engine benchmarked against industry simulators in prior work. The framework formalizes a semi-automated pipeline from Autodesk Revit through Dynamo BIM to COSMOS via a governed parameter store, converting BIM-derived quantity takeoffs into stochastic simulation inputs. This study makes two contributions: The first is the development of a governed BIM-to-simulation data pipeline in which an identifier-keyed data contract and in-flow validation establish an auditable chain of custody from design element to simulation input. The second is an advancement toward addressing a limitation the engine’s own validation studies identify—execution-phase site–data integration—through a control loop that couples Earned Value Analysis in a 4D BIM environment (Synchro) with iterative COSMOS-based scenario testing, enabling managers to evaluate corrective actions quantitatively before site implementation. Feasibility is demonstrated on an illustrative reinforced concrete building schedule, in which a schedule-performance shortfall triggers the loop and stochastic forecasting exposes an upper-tail completion risk hidden by the deterministic estimate. By establishing a governed pathway from BIM to the previously BIM-isolated COSMOS engine, this work provides a basis from which field-based evaluation of BIM-integrated stochastic project control can proceed. Full article

The rapid growth of express packaging waste (EPW) is creating increasing challenges for urban waste management and high-quality resource recovery, while reliable automated sorting remains difficult because of heterogeneous materials, deformable packaging, random stacking, and frequent partial occlusion. This study develops a system-oriented RGB-D sensing framework to improve the reliability of EPW sorting in conveyor-belt recycling operations. A synchronized and spatially registered RGB-D paired dataset, MEPWaste, was constructed containing 4528 paired samples (9056 RGB/depth images) from 10 representative categories of inner and outer packaging under multiple levels of stacking and occlusion. Based on this dataset, a multimodal detection framework was designed to improve feature reliability within each modality, reinforce spatial consistency between modalities, and enhance detection stability in densely stacked scenes with fragmented visible evidence. The proposed framework achieved 89.1% ${mAP}_{50}$ and 69.3% ${mAP}_{50:95}$, exceeding the RGB-only model by 5.3 and 5.8 percentage points, respectively, and the RGB-D baseline by 2.7 and 3.4 percentage points, with recall reaching 85.1%. A closed-loop sensing–inference–execution pipeline integrating synchronized RGB-D acquisition, real-time detection, robotic grasping, and PLC-based control was further implemented to examine engineering feasibility. Overall, the study provides a practical technical pathway for improving EPW sorting reliability and supporting intelligent resource recovery in waste management systems, with potential applicability to green campus management scenarios characterized by high-frequency parcel consumption and concentrated packaging waste generation. Full article

To support Industry 5.0 smart factories that require ultra-low latency and high reliability, this paper proposes a three-layer Intent-Based Zero-Touch Networking (IBZTN) pipeline. Existing Intent-Based Networking (IBN)/Zero-Touch Networking (ZTN) studies often remain conceptual, while Graph Neural Network (GNN)-based Quality of Service (QoS) prediction and Deep Reinforcement Learning (DRL)-based reconfiguration are usually developed as separate modules. The proposed pipeline connects natural-language intent translation, feasibility prediction, and minimal-modification reconfiguration through a validated QoS contract and feasibility-aware closed-loop structure. Layer 1 converts intents into quantitative QoS profiles by combining Retrieval-Augmented Generation (RAG) with schema- and rule-based validation. Layer 2 evaluates feasibility using a Graph Isomorphism Network with Edge features (GINE)-based binary classifier. Layer 3 recovers infeasible states using a Behavior Cloning (BC) Proximal Policy Optimization (PPO) agent with Smart Traffic Engineering (TE) masking. In experiments with 300 natural language intents, RAG+Validator reduced Layer 1 constraint violations to 0.0% for most evaluated cloud and local Large Language Models (LLMs). The Layer 2 predictor achieved a 93.9% F1-score, and Layer 3 achieved an 87.8% recovery success rate with 9.8 average modifications and 5.56 ms inference latency. These results demonstrate the simulation-level potential of IBZTN and motivate future hardware-in-the-loop validation in industrial networks. Full article

Artificial intelligence (AI) is transforming modern agriculture from experience-driven practices to data-driven production paradigms. To provide an in-depth analysis of AI technologies in intelligent agriculture, we retrieved literature from Web of Science, IEEE Xplore, Google Scholar and Scopus, covering publications from 2015 to 2025, and 85 articles remained after screening 1867 relevant publications. These articles are grouped into three stages from perception, to decision making, to execution (PDE) in a closed-loop framework. At the perception level, we highlight progress in intelligent sensing systems, such as unmanned aerial vehicle (UAV) and multi-modal monitoring platforms, for crop disease and pest detection, growth monitoring and abiotic stress assessment. At the decision making level, integration of heterogeneous data sources, including meteorological records, soil measurements, remote sensing (RS) imagery and market information, supports advanced analytics, such as yield prediction, pest and disease warning, irrigation and fertilization planning, and crop management optimization. At the execution level, agricultural robots equipped with simultaneous localization and mapping (SLAM) and deep reinforcement learning (RL) facilitate precision spraying, autonomous harvesting, and unmanned field operations. Overall, AI technologies demonstrate substantial potential in the PDE pipeline of agricultural production. However, several challenges remain, including heterogeneous data fusion, limited generalization across diverse environments, complex system integration, and high hardware and deployment costs. Future directions are discussed from the perspectives of lightweight model design, cross-platform standardization, enhanced human–machine collaboration, and a deeper integration of emerging AI paradigms to support scalable, robust, and autonomous agricultural intelligence systems. Full article

Image stitching is a computational technique designed to align and seamlessly fuse multiple overlapping images into a single panoramic image with an extended field of view. It plays a critical role in diverse domains, including mobile photography, autonomous navigation, and visual perception systems. However, most conventional image stitching pipelines implicitly assume that the input images have been pre-corrected for geometric distortions, particularly radial distortion inherent to wide-angle and fisheye lenses. This assumption often fails in practice, as many consumer-grade cameras lack built-in correction or calibration support. Consequently, applying standard image stitching methods to the uncorrected imagery frequently degrades feature correspondence reliability and introduces visible geometric misalignments and seam discontinuities in the final panorama. To overcome these limitations, this paper introduces a task-driven joint iterative optimization framework for image stitching that unifies unsupervised radial distortion correction, distortion-aware feature registration, and seam-aware blending within a single cohesive optimization objective. Specifically, lens distortion parameters are explicitly modeled as learnable variables and embedded into both the geometric registration and seam optimization sub-problems. An efficient closed-loop optimization strategy is then employed to jointly refine distortion parameters, homography estimates, and optimal seam paths in an alternating, mutually reinforcing manner. Implementation-wise, we first propose a calibration-free initial radial distortion estimation method which leverages intrinsic image gradients and epipolar consistency to provide physically plausible initialization for subsequent optimization. During iteration, distortion parameters are progressively refined by integrating robust geometric constraints derived from current feature matches (via RANSAC-based consensus filtering) with photometric consistency cues. Extensive experiments on multiple public benchmarks featuring pronounced radial distortion demonstrate that our method achieves superior stitching fidelity using metrics including PSNR and SSIM. It also confirms enhanced feature matching stability, which outperforms both distortion-agnostic approaches and two-stage pipelines that decouple distortion correction from registration. Furthermore, comprehensive ablation studies quantitatively and qualitatively validate the functional necessity and synergistic contribution of each core module, confirming the design rationale and effectiveness of the proposed joint optimization architecture. Full article

Advancing environmental monitoring through coordinated autonomous systems is central to sustainable smart region governance and data-driven territorial management. The article presents an engineering-oriented architecture and deployment methodology for an integrated wildfire monitoring and response system that combines unmanned aerial vehicles (UAVs), unmanned ground vehicles (UGVs), and stationary sensor networks (SNs). We formalise hub-and-spoke infrastructure placement as a mixed-integer optimisation problem that accounts for platform types, endurance, travel times and logistical constraints, and propose a practical pre-processing pipeline (confidence scoring, resampling, Kalman/median filtering, strategy fusion) for heterogeneous telemetry and imagery. The system couples multimodal neural network processing (image backbones, clustering and time-series models) with online resource-allocation and mission-planning mechanisms to prioritise UAV/UGV sorties and dynamically select launch sites. The article describes scenario-driven operational modes (early warning, alarm verification, autonomous local extinguishing, post-fire recovery, sensor-gap compensation, and inter-hub reinforcement), defines validation protocols (synthetic experiments, precision/recall/F1, and hardware-in-the-loop testing), and proposes KPIs to assess environmental, social, and economic impacts for smart regions. The contribution is a reproducible, deployment-focused blueprint that bridges conceptual UAV–UGV–SN research and practical implementation, highlighting trade-offs in reliability, communication redundancy, and sustainability, and outlining directions for simulation, field pilots and algorithmic refinement. Full article

The transition toward sustainable energy systems is increasing the operational complexity of modern power grids due to the high penetration of renewable energy sources, distributed energy resources, and bidirectional energy flows. Artificial intelligence has emerged as a key enabling technology for forecasting, optimization, and control in smart grids. Current AI implementations in energy systems lack unified workflows integrating forecasting, decision-making, adaptive stability regulation, and distributed coordination. Moreover, existing control approaches rarely incorporate biologically inspired stability mechanisms such as homeostatic regulation, limiting system-level resilience under dynamic operating conditions. This work aims to develop an architectural framework in the form of a unified artificial intelligence meta-pipeline enabling homeostatic control and distributed resource optimization in sustainable energy systems through closed-loop intelligent operation. A layered artificial intelligence meta-pipeline architecture is proposed integrating system representation, data intelligence, decision intelligence, homeostatic feedback regulation, and distributed coordination. A formal Homeostatic Energy Index is introduced to quantify system stress and enable supervisory adaptive policy regulation. The framework is validated using a reproducible microgrid-level simulation combining reinforcement learning-based control with homeostatic feedback regulation. Experimental validation demonstrates stable closed-loop operation under stochastic demand and renewable variability. The framework maintains bounded system stress levels, achieving an average Homeostatic Energy Index of 18.17 while preserving near-zero energy imbalance performance, confirming that homeostatic feedback improves stability without degrading energy balancing performance. This work introduces a unified artificial intelligence meta-pipeline architectural framework and formally defines a homeostatic feedback layer for sustainable energy system control. The proposed approach enables stability-aware structured integration of heterogeneous AI components and provides a foundation for self-adaptive, resilient, and distributed intelligent energy systems. Full article

To better understand the complexity of biological systems, research has shifted from a reductionist to a holistic approach, expanding the focus from single genes to a genome-scale view of gene activity and regulation. This is known as transcriptomics, a continuously growing field generating gene expression signatures from different technologies. A comparable paradigm shift has occurred in computational systems biology with the implementation of Artificial Intelligence (AI) learning models for gene expression analysis and integration. These models enable transcriptome-based profiling to address challenges of data heterogeneity, integration, and updating, assisting human intelligence and enhancing their ability to retrieve, analyze, integrate, and generate data recursively, thanks to their intrinsic predictive, inferential, reinforcement, and generative capabilities. Additionally, while scientists worldwide are still learning how to leverage AI methods that can maintain the human-in-the-loop, a new fundamental change is emerging: agentic AI, which can autonomously act and employ other AI methods to pursue its objectives. As a futuristic perspective, the proposed data analysis pipeline imagines agentic AI systems allowing the automated retrieval and pre-processing of heterogeneous transcriptomics data, analysis and integration with other omics datasets, performed with an incremental updating and recurrent analysis (IURA) model that could allow the detection of guideline updates (e.g., disease reclassification) and the generation of new hypotheses, such as candidate biomarkers or transcriptome–phenotype correlations. Since personalized medicine could derive profound benefits from its use, this scenario also raises important considerations regarding the advantages and concerns associated with the use of scientific AI agents in research and clinical practice. Full article

The rapid progress of large language models (LLMs) has fostered the development of domain-specific variants in law, medicine, and finance. However, existing legal LLMs still struggle to generate contextually grounded and regulation-compliant responses in complex scenarios of fair competition review. To address this, we present FairAgent, a collaborative multi-agent framework that unifies data refinement and reinforcement learning for legal reasoning. FairAgent integrates two core modules: (1) EchoCourt, a closed-loop data generation and refinement pipeline that constructs high-quality question–answer pairs through generation, critique, and optimization guided by a hierarchical Fairness Knowledge Forest; and (2) a two-stage outcome-based reinforcement learning mechanism that progressively teaches the model to invoke and integrate external retrieval in reasoning. We further enhance learning stability through a RAG-based rollout and retrieval-mask loss. Extensive evaluations demonstrate that FairAgent significantly improves reasoning accuracy, interpretability, and compliance in fair competition review compared with state-of-the-art baselines, establishing a scalable framework for retrieval-augmented legal intelligence. Full article