At open-pit mine intersections, autonomous haul trucks compete for temporary use of shared conflict regions under long braking distances, extended conflict-region occupancy times, load-dependent motion, and operational priorities. Existing intersection controllers are commonly tied to a fixed lane diagram, which obscures the invariant scheduling resource when benches, ramps, and haul roads are reconfigured. This study formulates the scheduling task as movement-conflict reservation (MCR) and develops the corresponding MCR-MARL method. Each configured junction is transformed into a directed-movement set and binary incompatibility matrix; runtime truck states generate predicted occupancy windows and current-step admission requests. The resulting MCR-MARL controller combines a load- and priority-aware mechanism reference, a parameter-sharing truck-local proximal policy optimisation actor that learns bounded residual accelerations, and a centralised deterministic reservation-feasibility operator that applies kinematic checks and priority-ordered compatible admission at every 1 s control step. The implementation is verified through structural checks, matched scheduling baselines, component ablations, and four-family leave-one-topology-out simulation. Across 12 held-out topology-seed combinations, MCR-MARL achieved 64.19 trucks per episode, 30.44 s average delay, and 11.27 s average stop time, improving over first-come, first-served in every paired comparison while matching strong rule-based schedulers. Relative to direct learning, the residual formulation increased throughput by 5.09 trucks per episode and reduced delay by 11.03 s. The truck-local actor used 84.0% fewer parameters and reduced decision latency by 46.6% relative to Ego-GAT. Removing conflict admission produced 2770 cross-movement events in 300 episodes, versus none under the full configuration. These results support a transparent simulation method for configuring and controlling the four evaluated mine-intersection layouts.
We audited execution fidelity and fault-classification agreement across three implementations of Benchmark Simulation Model No. 1. Reanalysis reproduced 174 archived metric rows, but actual sensor logs revealed default rather than registered noise in 59 MATLAB closed-loop runs. A separate 60-run correction verified 80,640 additive-noise values to a maximum error of 4.44×10⁻¹⁶. Registered-perturbation labels remained inconclusive, although the median MATLAB–QSDsan high-fault oxygen-trajectory RMS difference fell from 0.37145 to 0.00011239 mg O₂/L. After a two-block exploratory screen, a separately frozen internal validation generated 24 new input blocks at a stronger growth-rate perturbation. All 288 core trajectories passed execution and signal checks; 48 additional MATLAB trajectories reinstated the known noise-binding defect as controls. Nitrate was evaluable in 21/24 blocks: all implementations agreed on one positive, 20 negative and three inconclusive block labels. The primary joint endpoint nevertheless had 3/24 failures, with one-sided exact 95% upper bound 0.2923, exceeding the inherited 0.20 benchmark. Oxygen remained inconclusive in every block. Actual-signal auditing detected the known defect in 24/24 paired blocks, whereas file-identity and structural checks detected none; corrected controls had 0/24 flags. The binding change altered actuator trajectories but no paired loop labels. These results establish a reproducible execution audit and bounded internal validation, while showing that successful execution and label concordance do not establish adequate evaluability, simulator interchangeability or physical-plant validity.
Real-time prediction of tissue deformation is essential for interactive ultrasound-guided percutaneous surgery simulation, but high-fidelity physics-based models often exceed practical online computational limits. This paper presents a 3D Just-in-Time (JIT) modelling framework for real-time prediction of volumetric tissue deformation during ultrasound-guided percutaneous procedures. An offline historical database of needle insertion poses and corresponding volumetric deformation fields is generated using a simulation engine. During runtime, the proposed method retrieves a gated neighbourhood of geometrically similar insertions, computes curvature-aware constrained interpolation weights, and predicts the deformation field in a reduced output space obtained through principal component analysis (PCA). The full volumetric displacement field is then reconstructed and used to deform a baseline ultrasound volume for slice extraction. Across the tested grid resolutions, the proposed PCA-JIT framework achieved millisecond-scale prediction times and is up to 90 times faster than the simulation engine used to generate the database. Prediction errors remained low across all tested mesh resolutions, and the predicted fields enabled deformation-consistent B-mode slice rendering from arbitrarily posed imaging planes. The proposed framework provides an accurate and computationally efficient surrogate for volumetric tissue deformation prediction and is well suited for interactive surgical simulation, ultrasound-based training, and future image-guided intervention systems.
Discrete-event simulators coordinate state transitions through logical time, yet the runtime interpretation of a temporal control can diverge from its declared unit along conversion, clock, and event-scheduling paths. Without hardware measurements or golden outputs, parameter sweeps can leave incompatible interpretations exactly aliased. We present boundary-aware identification of temporal-control semantics, which reconstructs declaration-to-consumer paths, derives candidate responses, varies component period and global tick quantum independently, establishes signature separability, and observes a contrast-preserving boundary. In gem5, every nominal single-axis design retains an exact alias among seven stationary candidates, whereas 94.74% of 299,250 legal joint designs distinguish all seven candidates. The procedure reproduces every predefined classification in the controlled and source-hidden cases, classifies stock behavior and all eight controlled semantic deviations, and preserves distinct non-identification outcomes. A historical replay identifies Absolute edge semantics in eight pre-repair Access Map Pattern Matching (AMPM) settings and Relative cycles in eight repaired settings; repaired prefetch degree increases at 72 matched nonzero epochs. A 32-arm dynamic voltage and frequency scaling (DVFS) matrix uniquely identifies Schedule snapshot, with identical physical-time sequences across global tick quanta. Across gem5 and the independently developed Structural Simulation Toolkit (SST), all 104 checks outside the classification grid, comprising 824 matched-event contrasts, agree with prediction. Two 105-run scaling campaigns show that a fixed-record bounded probe remains at 2385 bytes and near-baseline wall time through one million events, whereas a full trace grows to 167 megabytes (MB) and 4.586–4.621× baseline. Together, these results provide a systematic and scalable basis for temporal-control semantic identification.
