Search Results (310)

Finite-control-set model predictive control provides excellent torque–speed regulation for switched reluctance motor drives but requires an online combinatorial search at every control instant, making low-cost embedded implementation challenging. This article investigates whether offline reinforcement learning can distill policies from multiple classical controllers into a single feedforward policy requiring neither online optimization nor controller gain tuning. A replay buffer is populated with trajectories generated by three demonstrators—hysteresis current control, proportional–integral control with pulse-width modulation, and finite-control-set model predictive control—using a finite-element model of a four-phase 8/6 switched reluctance machine parameterized from measurements of the physical drive. An implicit Q-learning agent then learns a control policy without evaluating actions outside the offline dataset. The central finding is that demonstration diversity governs the stability of offline reinforcement learning on this problem: policies trained from a single demonstrator experience early mode collapse in all fifteen runs, whereas two- or three-demonstrator datasets converge stably in all fifteen. Behavior cloning trained on the identical buffer, split, architecture, and deployed controller provides the reference point for interpreting this result. It matches the offline RL policy on torque quality and improves on its speed regulation, exhibiting none of the seed-to-seed fragility seen at no load while requiring roughly $8\%$ more switching transitions. The stability requirement therefore appears to be a property of the advantage-weighted offline RL objective rather than the control task, and the measured benefit of that objective on this problem is confined to switching effort. We report this rather than claim a broader advantage. The characterization of the distilled controller shows that it generalizes to operating points that are not included in the training dataset, gains nothing systematic beyond approximately 60% of the replay buffer, remains insensitive to $\pm 20\%$ perturbations of all reward weights, and degrades gracefully under measurement noise while the current mask enforces the peak-current constraint throughout. A deployment analysis shows that the 18,432 multiply–accumulate policy meets a $50\mathsf{\mu }$s control period in its existing form at a measured cost of about $2\%$ in torque ripple. All the results are simulation-based on a finite-element model parameterized from a physical machine. Full article

To address charging-resource contention and task-allocation conflicts in multi-UAV operations supported by a single mobile charging vehicle (MCV), this paper develops GEMS-DQN (Global-Enhanced Multi-Agent Scheduling Deep Q-Network), a centralized discrete-action scheduling framework for coordinated task and charging decisions. The framework uses a joint individual–global state representation to characterize UAV energy, task urgency, spatial information, global task progress, and charging-resource utilization. A normalized system-level reward with a dynamic conflict penalty provides explicit feedback for task-assignment and charging-resource conflicts. Per-UAV Q-values are used for feasibility masking and top-k action ranking, while beam search constructs a bounded joint-action candidate set for Monte Carlo Tree Search (MCTS) under stochastic MCV motion. Experiments are conducted over 30 independent training runs. At 800 training iterations, GEMS-DQN achieves a total score of $883.7\pm 22.4$, a task completion rate of $92.1\pm 3.4\%$, an average energy consumption of $10.3\pm 0.5\%$, and a conflict rate of $0.091\pm 0.018$. Compared with MAPPO, the strongest modern MARL baseline evaluated, GEMS-DQN improves total score by approximately $5.6\%$ and task completion by 7.9 percentage points, while reducing average energy consumption by 0.4 percentage points and conflict rate by 0.050. Ablation, reward-sensitivity, and scalability analyses further demonstrate the complementary effects of global information, conflict-aware learning, and bounded look-ahead search, while revealing the expected computation–performance trade-off of the centralized framework. Full article

Background: Breast imaging ends in a management decision, from routine return to biopsy, yet most models are evaluated by accuracy alone and say nothing about when they may be wrong. We present a trust-aware multimodal model that reads paired mammography and ultrasound, recommends one of three BI-RADS-derived management actions, and returns to the radiologist the cases it cannot call. Methods: Two foundation encoders are adapted with low-rank adapters and combined by a mask-aware fusion head. A trustworthiness layer adds calibration, conformal prediction sets, selective deferral, and an atypicality flag. We evaluated our method on the Breast Cancer Multimodal Imaging Dataset (BCMID), comprising 332 cases from 323 patients at a single centre. The reference standard is a three-class management grouping we derive from the reporting radiologist’s BI-RADS assessment, so performance is in concordance with that derived label rather than with pathology, observed patient management or longitudinal clinical outcome. Results: Macro AUROC was 0.759 and balanced accuracy was 0.556. Isotonic calibration reduced calibration error from 0.088 to 0.052, and prediction sets reached an empirical coverage of 0.934 at a mean set size of 2.32. Deferring the least confident 30% by a retrospective ranking of the pooled cohort raised balanced accuracy to 0.631. Two of 63 positive-management cases were under-triaged and 15 routed to additional imaging, with a recall of 0.730; freezing the encoders left macro AUROC at 0.756 but raised the under-triage count to twelve. Conclusions: Error rate and error direction are separable properties, and neither accuracy nor macro AUROC records the direction. The system therefore pairs each recommendation with calibrated probabilities, a conformal set of plausible actions and an explicit defer option, so uncertain cases return to the radiologist. These results establish an internally validated operating profile for radiologist-facing support; clinical safety, deployment readiness and benefit to patients still require external and prospective evaluation. Full article

Early-exit deep neural networks (DNNs) can reduce edge-inference latency, but abrupt variations in wireless and computing resources can destabilize split-inference policies. This paper proposes a meta-learning-driven adaptive control framework for joint backbone splitting and early-exit routing in MobileViT. The framework formulates multi-exit splitting as a constrained Markov decision process (CMDP) and introduces splitting-aware multi-dimensional adaptive proximal policy optimization (SMAPPO). SMAPPO combines nonlinear quality-of-service (QoS) penalties with topology-aware action masking, while cross-environment meta-initialization supports edge-local adaptation after resource disturbances. Under the stated simulation assumptions, SMAPPO reached the highest performance-index plateau among six methods in a representative 500-episode stationary trace and achieved the lowest normalized total cost across three latency–energy preference settings. Across ten seeds and nine stationary or disturbed scenarios, online SMAPPO achieved a 77.20% measured accuracy and 22.40 mJ of system energy. With an adaptation horizon of $K=14$, SMAPPO yielded a post-disturbance mean latency of 37.68 ms, a QoS-violation rate of 2.24%, and an on-time completion rate of 98.69%. These results indicate that combining meta-initialization, nonlinear constraint shaping, and topology-aware action masking improves stationary optimization and disturbance recovery within the controlled simulator. Full article

We present a reward-free control framework for balancing and steering a two-wheeled scooter using a diffusion-based world model. Rather than engineering a reward, we specify goals directly in observation space: target values (e.g., zero roll and zero yaw error) are pinned through a continuous mask, and classifier-free guidance amplifies the goal signal during trajectory generation. Because the mask is continuous at inference, goals can be traded off online (for instance, relaxing the balance constraint during sharp turns to allow necessary leaning) without retraining. The model is a FiLM-Mixer denoising network trained with V-prediction diffusion. At deployment, the controller runs in real time using a single diffusion step with warm-started predictions. We validate the approach on a full-sized Thormang3 humanoid operating a Gogoro Viva scooter in simulation, and deploy it on physical hardware. It matches a PPO baseline tuned with six reward components on balance, survival, and heading tracking while producing smoother commands, all without the per-task reward-shaping step. Diffusion training introduces its own loss-weight hyperparameters; unlike reward weights, however, these are task-agnostic. They govern the denoising procedure rather than the desired behavior, and are therefore set once and reused unchanged across goals rather than re-tuned for each new task. Because the model learns to predict trajectories rather than to maximize a reward, its training signal depends only on observed states and actions, not on reward labels. Real hardware recordings can therefore be folded directly into the same loss, providing a route toward closing the sim-to-real gap that reward-based methods such as PPO structurally cannot use. Full article

Large-scale public events generate concentrated passenger demand during egress periods, often overwhelming urban transit systems. This paper proposes a multi-modal evacuation framework that coordinates in-service buses temporarily diverted from existing lines and dedicated shuttle vehicles pre-positioned at depots. The problem is formulated as a two-layer stochastic optimization under travel time uncertainty: the upper layer determines pre-event shuttle fleet sizing, while the lower layer makes real-time dispatching decisions for both modes. We propose an Uncertainty-Aware Reinforcement Learning framework with Categorical DQN (UARL-CD) that learns a robust dispatching policy through a reward function aligned with the lower-level objective, explicitly accounting for travel time uncertainty via distributional value representation and stochastic training, with an action masking mechanism enforcing operational constraints. Simulation experiments based on a realistic stadium evacuation scenario demonstrate that the proposed framework significantly outperforms deterministic optimization and rule-based strategies, achieving a 31.6% reduction in evacuation completion time and a 48.4% reduction in average passenger waiting time compared to shuttles alone, while maintaining robustness to travel time uncertainty with only 4.0% performance degradation and online decisions executed within the 2-min decision interval. Full article

In disaster response and other infrastructure-limited settings, UAV-mounted access points can rapidly restore service availability for mobile ground users as demand and fleet availability evolve. Existing single-slot coverage formulations, however, can mask prolonged individual outages and do not jointly represent heterogeneous service priorities, finite battery capacities, and periodic recharging. We study persistent geometricmulti-UAV service coverage, where a user is available for service when it lies inside a UAV footprint. We propose Priority- and Outage-Guided Safe QMIX (POGS-QMIX), a hybrid hierarchical framework in which a centralized online coordinator forms conflict-reduced UAV–user targets from fleet-wide priority and outage information, while parameter-shared QMIX agents independently choose target-conditioned low-level actions. The framework couples class-balanced outage memory, assignment, dense target-progress feedback, and a return-energy action mask. The evaluation includes learning and non-learning baselines, greedy-versus-Hungarian assignment, multi-seed statistics, sensitivity studies, operating-condition studies, and energy-stress tests. In the default scenario, POGS-QMIX obtains high-priority coverage $0.547\pm 0.009$ and maximum high-priority outage $38.0\pm 4.7$ slots over five independent seeds. Full article

Advancements in intelligent manufacturing require solutions to the conventional flexible job-shop scheduling problem (FJSP) to accommodate increasingly intricate constraints, particularly in the semiconductor and electronic component sectors, where batch-processing machines (BPMs) significantly intensify scheduling complexity. To address this challenge, this study formulates an extended FJSP with multiple BPMs and proposes an end-to-end two-layer multi-agent deep reinforcement learning framework. Job and machine agents perform decentralized action mapping, while workshop states are encoded using a heterogeneous disjunctive graph and a dual-graph attention network. Unlike standard FJSP learning methods that primarily address operation–machine decisions, the proposed framework jointly models machine assignment, operation sequencing, variable-length batch formation, and BPM allocation within a unified policy, with a pointer network-based batching agent and an equipment-selection agent that handle batch-processing decisions under feasibility masking. The framework was validated using plant-derived production data and multi-scale synthetic instances. Numerical results show that the proposed method achieves competitive performance across the tested batching and standard-FJSP settings. In standard-FJSP comparisons, relative performance was scenario-dependent: DANIEL performed better in S1, whereas both proposed variants outperformed all comparators in S2. These results support the framework as an effective scheduling approach for deterministic FJSP with BPMs and indicate cross-scale generalization across evaluated instances. Full article

Barrier-based safety systems are fundamental to preventing accidents in high-hazard industrial operations. However, traditional Hazard Identification (HAZID) and risk screening frameworks aggregate safety safeguards at a macro-hazard level, creating a systemic blind spot that masks threat-specific vulnerabilities and single points of failure. To address this gap, this study develops ‘Maximum-Vulnerability Diagrams’ (MVD), a network-based modeling approach that maps and quantifies threat–barrier pathways using matrix algebra and conditional probability. Validated across empirical cases of working-at-height (H-06.01) and heavy rotary equipment (H-08.01) operations in the oil extraction industry, the MVD successfully isolates high-criticality, zero-redundancy pathways. We mathematically establish that multiplexed defenses require a target individual efficiency of η ≥ 95% to reliably suppress system failure probability below a strict 5% operational threshold. The findings demonstrate that aggregate safeguard volume is a deceptive safety metric, and that systemic resilience depends entirely on network architecture. This framework transitions risk governance from passive compliance checking to predictive, threat-driven barrier management, offering an actionable methodology to optimize safety resources before accidents occur. Full article

To address the challenge of satisfying strict Service-Level Agreement (SLA) requirements for concurrent smart video surveillance tasks in heterogeneous edge computing environments, an SLA-aware adaptive scheduling algorithm for heterogeneous computing collaboration is proposed. First, a mixed-task flow model is constructed, and a finite-state Markov chain is utilized to dynamically model the time-varying wireless channel. Second, a Dueling Double Deep Q-Network (Dueling DDQN) scheduling algorithm based on SLA awareness and channel adaptation is proposed, with a designed SLA action-masking mechanism. This mechanism advances hard delay constraints to the decision-generation stage, dynamically prunes the action space based on real-time channel conditions and node loads, and filters out actions predicted to violate the SLA before execution. Experimental results show that the proposed algorithm coordinates heterogeneous computing resources between the cloud center and the edge and exhibits earlier empirical reward stabilization and lower task-violation rates than the compared learning-based baselines under the tested workload conditions. 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

Randomized controlled trials (RCTs) are often considered the gold standard for causal inference, but their implementation can be costly, time-consuming, and sometimes infeasible due to ethical or practical constraints. The so-called target trial emulation framework introduced the systematic use of observational data for treatment effect. This approach necessitates the detailed specification of a hypothetical trial protocol including eligibility criteria, treatment strategies, and outcome measures, which are then emulated by utilizing observational data. We expanded the target trial framework by integrating drug pathway embeddings and causal modeling, enabling prediction of treatment outcomes for unseen or held-out mechanisms of action based on the embedding relationships among existing therapies. We demonstrate that embedding-based models can reliably predict the direction of observed clinical outcomes across diverse therapeutic classes (e.g., small molecules, biologics), even when masking the observational data for the particular mechanism being estimated, though the precise magnitude of treatment effect remains hard to recover. This approach illustrates the potential to estimate the clinical efficacy of new drug mechanisms and to enhance the precision of future trial design and operations. Full article

Background: Predictive performance does not establish whether model-identified risks correspond to feasible, equitable pathways. We assessed the actionability and equity of a severe tooth loss prediction model using constrained algorithmic recourse and same-source temporal validation. Methods: An Explainable Boosting Machine was trained using 2022 Behavioral Risk Factor Surveillance System data from 433,772 adults. High-risk adults underwent recourse auditing incorporating immutability locks, behavioral directionality, physiological safety floors, and at most three feature changes. Recourse represented hypothetical movement within model space, not treatment advice, causal risk reduction, or reversal of tooth loss. Primary reachability was the unweighted analytic-cohort proportion for which the frozen engine identified a feasible pathway; a BRFSS survey-weighted domain sensitivity analysis used final weights, strata, and primary sampling units. The audit separated reachability from conditional burden among reachable adults. Equity was evaluated across Social Indicators of Disparity Index (SIDI) and income strata using Oaxaca–Blinder decomposition. The frozen specification was evaluated in a 2022 holdout and the 2024 BRFSS cohort (N = 448,213) without retraining. Results: Survey-weighted areas under the receiver operating characteristic curve were 0.858, 0.855, and 0.858 in the 2022 full, 2022 holdout, and 2024 cohorts. Primary unweighted reachability was 10.61%, 10.54%, and 9.82%; corresponding survey-weighted estimates were 12.69% (95% design-aware CI, 12.33–13.07%), 12.31% (11.54–13.12%), and 11.30% (10.95–11.65%), respectively. Thus, reachability remained limited under both estimands. The high-SIDI group showed poorer calibration. Under the prespecified primary SIDI-neutral additive-cost specification, residual conditional cost differences across SIDI strata were small; alternative burden metrics were direction-dependent. The income residual attenuated from the 2022 holdout to 2024, although its confidence intervals were fixed-pipeline row-bootstrap intervals rather than fully design-based intervals. Conclusions: Stable predictive performance masked limited model-space actionability. Integrating explicitly labeled reachability estimands, conditional burden, equity, and temporal transport can strengthen pre-deployment evaluation. Because severe tooth loss is irreversible, recourse pathways should be interpreted as a stress test of model-implied modifiable factors rather than evidence of reversibility. Full article

Scheduling at the printing bottleneck of plastic woven packaging production is complicated by asymmetric sequence-dependent setup times (SDST) caused by color transitions and by mandatory non-preemptive shutdown windows. This study focuses on assigning and sequencing orders on parallel printing machines under these coupled operational constraints. To address this problem, we propose an Attention-Enhanced Dual-Critic Deep Reinforcement Learning framework with Action Masking (AE-DMAC). The scheduling problem is formulated as a bi-objective parallel-machine Markov Decision Process that considers makespan and total setup time under a fixed preference setting. A cross-attention module models the compatibility between the current machine state and pending orders to capture asymmetric SDST effects. Two critics separately estimate the efficiency- and setup-related value signals before they are combined for policy optimization, reducing interference between the two objectives. In addition, a deterministic action-feasibility mask removes assignments that would overlap the known shutdown window before action sampling. The framework is evaluated on industrially calibrated synthetic instances with 50, 150, and 300 orders scheduled on eight parallel printing machines. Experimental results show that AE-DMAC consistently improves makespan and normalized setup time per machine compared with the implemented heuristic, meta-heuristic, and Vanilla PPO baselines under the tested operating conditions. In the large-scale instance, AE-DMAC achieves an average makespan of 223.8 h and a normalized setup time of 568.4 min per machine, corresponding to reductions of 7.4% and 27.4%, respectively, relative to Vanilla PPO. The feasibility mask maintains zero shutdown-window violations in the evaluated deterministic setting, while the attention mechanism substantially reduces high-cost sequence-dependent color transitions. These results indicate that the proposed framework is a promising scheduling approach for parallel printing systems with asymmetric changeovers and known machine-unavailability windows. Full article

The implementation of the Air Quality Health Index (AQHI) is a regional climate action intended to protect population health, but its empirical linkage to the sustainability goal of good health and well-being remains untested. Using mortality data from Hong Kong (2013–2022), we employed a two-month-lag fixed-effects model to estimate the change in years of life lost (YLL) associated with monthly AQHI warning frequency. We further assessed heterogeneity across demographic and geographic subgroups, and explored the underlying mechanisms. The findings show that monthly AQHI warning frequency is negatively associated with YLL two months later. Each additional warning per month is associated with a reduction of 0.0032 units (SE = 0.0011) in per capita YLL across districts in Hong Kong. The baseline result was further supported by extended survival analyses. Heterogeneity analyses showed that each additional AQHI warning was associated with a significant reduction in YLL among males (coef. = −0.0066, SE = 0.0001) and those with spouses (coef. = −0.0065, SE = 0.0083), but not among females or those without spouses. Mechanism analyses suggested that the effect of AQHI warning frequency on YLL reduction was significantly moderated by individual behavioral responses, including increased face mask usage (int. coef. = −0.0047, SE = 0.0016) and reduced short-term travel (int. coef.= −0.0069, SE = 0.0004), and through socio-environmental pathways such as reduced traffic accidents (int. coef. = −0.0134, SE = 0.0034) and lower carbon emissions (int. coef. = −0.0367, SE = 0.0112). Hong Kong’s AQHI forecasting experience demonstrates the viability of health risk warnings as a climate adaptation strategy in sustainable urban governance, while its observed equity gaps offer critical lessons for refining inclusive environmental health policies. Full article