With the rapid advancement of autonomous driving and navigation technologies, an increasing number of vehicles rely on real-time routing services such as Google Maps and Waze. However, these services frequently fail to improve aggregate traffic conditions, as rerouting drivers based on existing congestion amplifies herding behavior and produces new secondary bottlenecks. To address this issue, a digital twin-based dynamic route-planning framework is proposed, in which real-time vehicle detection, surrogate-modeled traffic-flow prediction, and a congestion-aware metaheuristic router are integrated into a closed-loop pipeline. A network of intersection cameras with a YOLOv8 detector supplies real-time inflow and outflow data, which is fed into a Lighthill–Whitham–Richards (LWR)-based simulator incorporating signal, lane, and weather factors. The counting stage is validated at the level of individual crossing events on continuously annotated traffic sequences. To deliver real-time predictions at network scale, a parallel partial Gaussian stochastic process (PP-GaSP) surrogate is adapted to edge-level traffic volumes, providing sub-millisecond predictions with closed-form Student-t uncertainty that the router consumes as a risk-inflated flow estimate. Since the single-vehicle social-cost objective is edge-separable and solved exactly by a modified Dijkstra search, the paper addresses routing where herding actually arises, jointly optimizing a batch of concurrent demand streams whose coupled objective contains a non-separable network-bottleneck term beyond the reach of per-vehicle best-response methods. A new metaheuristic, Adaptive Chaotic Bacterial Foraging Optimization (ACBFO) is proposed to solve the batch problem through congestion-guided chemotaxis, which reroutes demand streams off the current fleet bottleneck within a chaotically scheduled bacterial foraging search. In paired experiments, ACBFO attains the provably optimal single-vehicle solution on every instance and significantly outperforms exact sequential routing, iterated best response, and metaheuristic and learning-based baselines, with the largest gains on the fleet bottleneck ratio, and a multi-window closed-loop replay exercises the prediction, joint-routing, and feedback stages under time-varying boundary-count inputs.
Efficient decentralized task scheduling remains a fundamental challenge in dynamic and heterogeneous edge computing environments, particularly for latency-sensitive and distributed applications. Existing approaches typically address reactive scheduling, resource prediction, or intelligent reasoning in isolation, limiting adaptability and robustness in real-world deployments. This paper presents PRECOG-Edge, a decentralized predictive–cognitive scheduling system that advances edge scheduling from reactive decision-making to predictive, cognitive, and autonomous multi-agent coordination. Unlike existing methods that treat prediction, reasoning, and coordination as separate components, PRECOG-Edge integrates resource forecasting, fairness-aware task placement, constrained Large Language Models(LLM)-assisted reasoning, and decentralized coordination through a unified predictive–cognitive scheduling pipeline operating on CPU-only edge devices without centralized control or global state. Lightweight Long Short-Term Memory (LSTM) models forecast short-horizon resource availability, and the resulting predictions are transformed into deterministic resource-aware suitability scores that guide decentralized task negotiation. A constrained LLM provides contextual and fairness-aware recommendations for candidate task allocations, while deterministic scheduling policies retain final authority over task allocation. The proposed system is evaluated on a four-node heterogeneous edge cluster across four operational modes. Experimental results demonstrate that predictive scheduling improves load stability and reduces hotspot formation relative to reactive scheduling, while cognitive scheduling enhances fairness-aware task allocation and context-sensitive decision support. In fully autonomous operation, PRECOG-Edge reduces load imbalance from an estimated Gini coefficient of 0.30–0.35 to 0.057–0.151 across 66 scheduling rounds while preserving decentralized coordination. These findings demonstrate that PRECOG-Edge enables interpretable, resilient, and autonomous task scheduling in heterogeneous edge environments.
Lightweight deep learning models are increasingly used for real-time crop disease diagnosis in resource-constrained agricultural environments. Although these models achieve high predictive performance with low computational cost, their black-box nature limits transparency, trustworthiness, and practical deployment. Existing explainable artificial intelligence (XAI) approaches are primarily used for post hoc interpretation, with limited utilization for guiding model improvement.This study proposes an explainability-guided optimization framework for lightweight wheat disease classification that elevates explainability from a passive interpretation mechanism into a structured, quantitative source of guidance for model enhancement within an expert-in-the-loop design process. The framework follows an explain–evaluate–diagnose–improve paradigm in which Grad-CAM, Score-CAM, and LIME are used to analyze model behavior, quantitatively evaluate explanation quality, and identify critical failure modes, including attention misalignment, background interference, weak disease localization, and inter-class confusion. The resulting insights are used to guide targeted architectural refinement through the integration of a lightweight XAI-guided attention module.Experimental results show that the proposed framework improves classification accuracy from 95.67% to 98.33%, while consistently increasing precision, recall, and F1-score across all disease categories. Quantitative explanation evaluation further demonstrates improvements in localization quality, background suppression, explanation stability, and failure diagnosis capability. Multi-run statistical validation confirms that the observed gains are reproducible and statistically significant, while ablation analysis shows that the improvements arise from explainability-guided refinement rather than attention integration alone. These benefits are achieved with only a marginal increase in computational complexity, preserving suitability for edge deployment. The improvements further transfer to an independent public wheat disease dataset, are confirmed under five-fold cross-validation and against standard attention modules (SE, ECA, CBAM) and widely used lightweight architectures, and are substantiated by measured on-device inference latency on a smartphone-class edge platform.The findings demonstrate that explainability can serve not only as an interpretability tool but also as a systematic and reproducible source of optimization guidance for developing accurate, transparent, and computationally efficient intelligent systems for precision agriculture.
In this study, a Quantum-Augmented Salp Swarm Algorithm with Adaptive Parameter Control (QSSA-APC) is presented as a novel optimization framework for effective hyperparameter tuning of Hybrid-CNN models in practical applications. The suggested method improves the convergence and optimization quality of the current Salp swarm algorithm (SSA) by balancing exploration and exploitation through the combination of quantum-inspired exploration and adaptive parameter control. Benchmark functions from IEEE CEC 2017, CEC 2020, and CEC 2022 under 50, 100, 10 and 20-dimensional problem spaces which span over unimodal, multimodal, hybrid, and composite issue classes are used to thoroughly assess the approach. Statistical significance is confirmed using the Wilcoxon signed-rank, Friedman and Holm post-hoc tests, supported by Cliff’s delta effect size analysis. Typical optimization measures, and convergence behavior are all part of performance analysis. Additionally, across a number of real-world image datasets, the suggested hybrid CNN–Transformer model modified by QSSA-APC shows appreciable gains in performance metrics and convergence analysis. The efficiency, higher robustness and scalability of QSSA-APC as a dependable metaheuristic optimizer for intricate, high-dimensional learning and engineering optimization problems are demonstrated by these results which proves it as a potential tool for realistic hybrid CNN hyperparameter optimization in challenging tasks.
