Many insects harbour microbial communities that can profoundly influence the biology of their host. Yet, the relative contribution of random exposure (i.e., stochastic) events and deterministic ecological factors in shaping these communities remains unclear for most taxa. We examined microbiome assembly across 344 firefly (Coleoptera: Lampyridae) specimens from the Northeastern United States, spanning 12 species and species groups, and generating a high-resolution dataset through deep 16S rRNA gene amplicon sequencing and quantitative PCR. To formally assess the balance between stochastic and deterministic forces, we applied integrative statistical approaches, including an innovative null-modelling framework based on the normalized stochasticity ratio (NST) index. We hypothesized that firefly microbiome assembly is dominated by stochastic processes driven by unpredictable microbial exposures. Consistent with this, we observed elevated NST values for most bacteria, coupled with high intraspecific variability in bacterial abundance and composition. However, microbiomes were more similar among closely related fireflies and unusually prevalent mollicute strains showed low NST values, species-specific associations and retention across geography and host development. While adult bioluminescence and diet could not be directly linked to microbiome abundance or composition, considering seasonal factors and intra-host anatomy within host species revealed patterns explaining some of the intraspecific microbiome variation. These results show that deterministic processes, likely arising from host-specific microbial filtering mechanisms, act alongside stochastic forces to shape firefly-microbe associations. By integrating broad field sampling with quantitative bacterial load estimates and comprehensive microbiome analyses, this study clarifies how evolutionary history, ecology and chance jointly govern microbiome assembly in a diverse insect lineage.
Hemorrhagic shock is a common emergency that accounts for more than 10% of global mortality and up to 40% of trauma-related mortality. The Advanced Trauma Life Support (ATLS) guidelines outline resuscitation strategies in patients with massive hemorrhage based on clinical trial and observational data. Early intervention with fluids and/or blood products is recommended during resuscitation as key to maintaining vascular patency, stabilizing a patient's blood pressure, maintaining tissue oxygenation, and limiting shock. From a transfusion perspective, there is wide heterogeneity in the quality of blood products due to donor and manufacturing variation. The optimal transfusion strategy for massive hemorrhage remains unclear. In silico models of resuscitation may offer a means to evaluate the effectiveness and safety of various resuscitation protocols. Here we describe a stochastic multicompartment model of fluid balance and resuscitation that includes (i) cardiovascular hemodynamics, (ii) body fluid compartments and capillary solute exchange, (iii) the ability to alter hemorrhage, resuscitation, and hemostasis parameters, and (iv) tissue oxygenation and capillary-alveolar gas exchange. Building upon deterministic frameworks, this stochastic model more faithfully reflects the clinical heterogeneity of bleeding patients at a level I trauma center. This allows for a proof of concept in silico trial comparing crystalloids, conventional component therapy (CCT), i.e. red cell, platelet and plasma components, to cold-stored low titer group O whole blood (LTOWB). In an in silico cohort of ATLS class III (>30% and ≤40% blood volume lost) hemorrhage, LTOWB resuscitation reduced the time spent in the critical hemostatic window (platelet count <50x109/L, INR≥2, hemoglobin (Hgb) <8 g/dL and fibrinogen <150 mg/dL) compared with CCT (90.59 vs. 147.62 minutes; p=0.04), with no difference in predicted event-free survival (t = 240 minutes; cardiac < 1.5, fluid overload > 10% or SBP > 150% of the starting SBP, or Hgb < 3.0 g/dL). In a separate cohort of ATLS class IV (>40% blood volume lost) hemorrhage, LTOWB yielded a higher predicted event-free survival (74%) versus the CCT arm (69%, p < 0.01). We demonstrate how this in silico platform can function as a digital twin for hemorrhagic trauma enabling precision transfusion strategies, and in parallel, an operational twin to support blood banks to forecast blood product demand and inform massive transfusion protocols and clinical trial design.
This article investigates a predefined-time tracking control problem for stochastic nonlinear multiagent systems (MASs). A novel continuous-time hysteretic quantizer is proposed for the first time, achieving seamless transitions between quantization levels. Controller chattering is eliminated while an arbitrary configuration of level-specific dwell durations is permitted. A deferred activation function is introduced to remove the dependence on initial state values in the prescribed performance funnel control design. The proposed method guarantees semi-global practical predefined-time stability for all closed-loop signals, with satisfactory tracking performance of the desired signal. The efficacy and robustness of the proposed approach in meeting prescribed performance specifications are validated through numerical simulations.
Deep brain stimulation (DBS) is a widely used therapy for neurologic and psychiatric disorders. Conventional DBS delivers highly regular stimulation patterns that suppress pathological activity but can induce stimulation-related side effects, limiting the therapeutic window. Introducing controlled temporal variability through stochastic pulse timing may represent an alternative programming dimension to improve tolerability while preserving clinical benefit. An adult in their 60's with bilateral Vim DBS underwent evaluation of tonic, pink-noise, and white-noise stimulation patterns delivered through his chronically implanted Boston Scientific Genus system using the Chronos research platform. We assessed tremor and stimulation-induced side effects using accelerometry, spiral drawing tasks, standardized speech recordings, and patient-reported paresthesias. Pink noise stimulation preserved meaningful tremor suppression while improving tolerability compared with conventional tonic 130 Hz stimulation. Under tonic stimulation, dysarthria and paresthesias were prominent at 2.0 mA, narrowing the usable therapeutic window. In contrast, pink noise maintained tremor control across the same amplitude range with reduced side-effect burden. White noise stimulation demonstrated intermediate effects, providing improved tolerability relative to tonic stimulation but less tremor suppression than pink noise. Findings were consistent across accelerometry and functional drawing tasks. This study provides first-in-human evidence that temporally structured stochastic pulse timing can preserve therapeutic benefit while expanding the tolerable stimulation range relative to tonic DBS. These findings suggest that temporal structure represents a clinically meaningful programming dimension that may broaden the DBS therapeutic window using software based updates to existing hardware. Further evaluation in larger cohorts is warranted.
We investigate the dynamics of an inertial active Ornstein-Uhlenbeck (OU) particle in the presence of stochastic resetting. Using the renewal approach, we compute the mean square displacement (MSD) and position probability distribution functions both in the overdamped and underdamped regimes. The impact of resetting and activity is significant only in the steady state. The steady state MSD gets suppressed with an increase in the resetting rate and shows a nonmonotonic impact on the duration of activity. For a very short and very long persistent duration of activity, the steady-state MSD is same as that of a passive Brownian particle. However, for an intermediate range of activity time, the steady state MSD is enhanced and shows a maximum, allowing the particle to explore a larger region, thereby increasing the probability of encountering a wide-spread target. These results are further supported by the position probability distribution function. Moreover, when the particle is suspended in a viscoelastic medium characterized by the presence of memory, the MSD interestingly develops an intermediate-time plateau and the plateau depends on the strength of viscoelasticity and viscoelastic relaxation time. An increase in viscoelasticity strength broadens the plateau and suppresses MSD. On the other hand, slower viscoelastic relaxation makes memory effects weaker, resulting in the disappearance of the plateau. Similarly, fast resetting dominates the memory effects and makes the system approach steady state faster, overcoming the intermediate plateau in MSD. Moreover, a longer relaxation time amplifies the enhancement of steady state MSD due to activity, making it a favorable condition for reaching the wide-spread target. Finally, our analytical results agree well with numerical simulation.
During a global pandemic, hospitals face challenges of uncertain patient influx and increased risk of absenteeism among medical personnel, both of which adversely impact patient safety. To address these challenges, we propose a two-stage stochastic program for nurse staffing that incorporates uncertainties in patient demand and absenteeism. Our model supports two critical staffing decisions to optimize nurse allocation. The first is a tactical decision regarding the number of nurses to be cross-trained, and the second is an operational decision concerning the number of temporary nurses that need to be hired. Applied to data from a Norwegian tertiary public hospital during the first wave of the COVID-19 pandemic, our model identifies a bottleneck in intensive care unit nurse availability. Sensitivity analyses reveal that the effect of increasing the penalty for untreated patients is much larger than the effect of changes in cross-training parameters, such as the number of trainees per mentoring nurse or cross-training cost. Moreover, we highlight that cross-training helps to reduce bottlenecks and improves future service levels. Thus, cross-training remains advantageous overall, despite temporarily reducing nurse availability during the cross-training period. Despite the study's limited scope on a single patient pathway and its focus solely on nurses, it provides valuable insights into nurse staffing strategies for practitioners. To the best of our knowledge, this is the first study to model cross-training as a tactical staffing decision and its implications for workforce availability during the cross-training period, while also accounting for an increased risk of absenteeism.
Scoring rules are critical for evaluating the predictive performance of epidemic models by quantifying how well their projections and forecasts align with observed data. In this study, we introduce the energy score as a performance metric for stochastic trajectory-based epidemic models. As a multivariate extension of the continuous ranked probability score (CRPS), the energy score provides a single, unified measure for time-series predictions. It evaluates both calibration and sharpness by considering the distances between individual trajectories and observed data, as well as the inter-trajectory variability. We provide an overview of how the energy score can be applied to assess both scenario projections and forecasts in this format, with a particular focus on a detailed analysis of the Scenario Modeling Hub results for the 2023-2024 influenza season. By comparing the energy score to the widely used weighted interval score (WIS), we demonstrate its utility as a tool for evaluating epidemic models, especially in scenarios requiring integration of predictions across multiple target outcomes into a single, interpretable metric.
Traditional psychological models often treat motivation as a predictable, linear process, failing to capture how religious drive shifts erratically between fluid detachment and intense fervor. To resolve this limitation, a generative nonlinear dynamical systems framework is developed wherein desire is formalized as a vector field defined by two macroscopic order parameters: motivational intensity and semantic orientation. A coupled system of stochastic differential equations models how religious practices manipulate control parameters, specifically systemic gain and landscape curvature. Numerical simulations utilizing Euler-Maruyama integration reveal that these parameter adjustments induce critical bifurcations between two distinct topological regimes. The first, characteristic of monotheistic strategies, generates a stable fixed point exhibiting cusp-like hysteresis, which protects belief systems against environmental noise. The second, characteristic of Daoist strategies, generates a metastable regime of flexible wandering, maximizing cognitive adaptability. These findings suggest that religious traditions function as dynamic control mechanisms for managing the stability-plasticity dilemma. Ultimately, the model bridges abstract phenomenology with empirical psychology, providing testable blueprints based on the cusp catastrophe model to evaluate sudden phase transitions and cognitive hysteresis in motivational regulation.
This article focuses on the self-triggered estimator for nonlinear Markov jump systems (MJSs) subject to stochastic hybrid attacks, including deception attacks (DAs) and denial-of-service (DoS) attacks. To alleviate communication pressure and eliminate the need for continuous detection, a dynamic self-triggered mechanism (DSTM) is proposed, wherein the dynamic variable is adaptively adjusted to conserve network resources more efficiently. Considering the inherent openness of communication networks and their susceptibility to attacks, two Markov chains are employed to characterize the random switching between different attack strategies. With the help of a mapping technique, the system modes and attack modes are integrated into a new Markov chain. In addition, a hidden Markov model (HMM) with incomplete transmission probabilities is considered to represent the mismatched modes and the covert characteristic of attacks. Then, sufficient conditions ensuring system stability are derived, and a dissipative estimator is developed. Finally, two examples are utilized to illustrate the efficacy of the derived theoretical results.
This paper addresses the poor stability of near-ground flexible capture nets under uncertainty in the ambient wind. The capture net considered is a near-ground system for intercepting small low-altitude targets such as small UAVs or individuals; given its small scale (a 3.2 m net) and short deployment time (about 2 s), the wind can be treated as a quasi-steady flow during a single deployment, and the relevant uncertainty is characterized through the variability of the near-ground ambient mean wind speed. Focusing on the factors influencing the flight attitude of capture nets, a performance evaluation index model for flexible capture nets is established, enabling quantitative assessment of net performance, and a near-ground flexible capture net dynamics model is constructed. The Morris-Sobol method is employed to perform global sensitivity analysis on the factors affecting the effective range of flexible capture nets, revealing that environmental wind speed, ballistic mass, launch angle, and cord diameter are the primary factors influencing capture net performance. The Pareto-optimal non-dominated solution set is obtained using a Gaussian Process Regression (GPR) surrogate model solved via the Non-dominated Sorting Genetic Algorithm II (NSGA-II), followed by objective decision-making based on the Utopia point method. The robust design optimum is determined as m_b = 22.34 g, d = 0.370 mm, and theta = 28.9 deg. Finally, high-fidelity numerical simulations are conducted for verification. Compared with the empirical design parameters, the mean effective range increases by 9.02% (statistically significant), and the standard deviation of the deterministic design is 45.71% higher than that of the robust design.
In this work, we revisit the Wells-Riley model, which has been widely used to estimate airborne infection risk in indoor settings. In particular, we consider a probabilistic (i.e., "stochastic") framework of the Wells-Riley model which allows one to quantify infection risk in terms of the per-capita probability of infection for each susceptible individual, as well as the probability distribution of the number of infections (here referred to as "exposures") during the indoor interaction. Directly extending the work by Edwards, King, Noakes et al. (2024), we consider here the situation where the main parameters in the Wells-Riley model (namely, the quanta generation rate q $q$ , the ventilation rate Q $Q$ , the number of infectors I $I$ , or the duration of the indoor interaction T $T$ ) may be random or uncertain. We show how, in this case, the per-capita infection risk P i n f e c t i o n $P_{infection}$ becomes a random variable between 0 and 1, and compute its density function under some parametric assumptions. This allows for a comprehensive analytical quantification of uncertainty when dealing with heterogeneous populations, uncertain environmental conditions, or stochastic human behavior. Our results reveal that infection risk can vary significantly depending on the distribution and variability of model parameters. In particular, using mean parameter values in the classical Wells-Riley model can lead to systematic inaccuracies: Uncertainty in q $q$ , T $T$ , or b $b$ leads to infection risk overestimation, while environmental stochasticity (i.e., uncertainty in ventilation or removal rates) can lead to infection risk underestimation. We also investigate which parameter mainly drives the uncertainty in infection risk when two model parameters are simultaneously random.
Voltage-gated ion channels play important roles in many membrane-enclosed structures, including synaptic vesicles, endosomes, mitochondria, chloroplasts, viruses, and bacteria. Here, we study how compartment size and channel gating interact to shape voltage dynamics and ion content in sub-micron structures. In small compartments, assumptions underlying conductance-based (Hodgkin-Huxley type) models of membrane voltage must be relaxed: (1) stochastic gating of individual ion channels can quickly and substantially change membrane voltage; (2) these changes can equilibrate faster than channel state dwell times; and (3) ionic currents, even though as few as two channels, can substantially alter ionic concentrations. We adapted conductance-based models to incorporate these effects, and we then simulated voltage dynamics of small vesicles as a function of vesicle radius and channel density. We identified regimes in this parameter space with qualitatively distinct dynamics. We then performed stochastic simulations to explore the role of NaV1.5 in the maturation of macrophage endosomes. The stochastic model predicted dramatically different dynamics compared with a deterministic approach. Electrophysiology of nanoscale structures can be very different from larger structures, even when ion channel composition and density are preserved.
Resistive memory (RM)-based computing-in-memory (CiM) accelerators provide a promising platform for low-power edge intelligence. However, practical edge AI requires not only efficient inference but also repeated model updates through class-incremental learning (CiL). Implementing CiL on RM substrates exposes a critical material-algorithm mismatch: the write-intensive updates required by CiL are strongly affected by the stochastic programming of filamentary RM devices. Conventional fully programming (FP) mitigates this stochasticity by repeatedly programming and verifying each cell to a precise target conductance, but this exhaustive procedure incurs substantial energy consumption, latency, and device wear across successive CiL stages. Herein, we propose a hardware-aware adaptive programming (AP) strategy that aligns CiL deployment with RM device physics. Microstructural and electrical analyzes reveal that the random spatial distribution of oxygen vacancies gives rise to unavoidable programming variability. Guided by this insight, AP does not attempt to eliminate intrinsic stochasticity through costly compensation. Instead, it updates only the most impactful weights to suppress accuracy loss caused by overall mapping errors, while leaving low-impact weights unchanged. This converts large-scale write-verify operations into targeted updates of a minimal subset of cells, reducing programming overhead without requiring device or material optimization. Validated on a hybrid analog-digital system with a 40 nm, 256 k RM-based CiM core, AP reduces programming energy by 93.0% and programming cycles by more than 90% relative to FP during five-stage CIFAR100 CiL, while achieving a final accuracy of 0.80, close to the 0.81 software baseline. For the more complex ShapeNet 3D point-cloud recognition task across eight stages, AP achieves 92.3% energy savings with only a 0.03 accuracy loss. Moreover, AP improves robustness by reducing programming-error-induced accuracy degradation by 87.7% and 88.4% for the two tasks, respectively. This work bridges algorithmic update requirements and physical programming constraints, enabling robust and energy-efficient lifelong learning on RM-based CiM platforms.
Current translational approaches to obesity and metabolic syndrome predominantly rely on chemical interventions modeled via stochastic passive diffusion. These traditional methods inherently lack spatial targeting and temporal precision, often resulting in delayed therapeutic onset. Here, we present a paradigm shift using a novel quantum biophysical framework that treats the gastrointestinal microenvironment as a non-equilibrium semiconductor network. By engineering a Metabolite Conjugate Prebiotic (MCP) platform with an embedded polyphenolic tannin matrix, we command targeted electron translocation across microbial membranes, establishing a directional "metabolic pull." Validated against a Real-World Evidence (RWE) cohort of 2,000 subjects, this platform demonstrated an accelerated biological response window, achieving rapid and deterministic activation of satiety signaling within 15 to 60 min. Notably, while the FDA recommends 20 to 25 grams per day for traditional prebiotics, participants in this cohort required a significantly minimized dosage of only 2 to 3 grams of MCP per day to elicit these homeostatic effects. The computational framework yields a 92 percent clinical predictive accuracy, significantly outperforming classical stochastic models. This platform offers a scalable, highly predictable avenue for precision translational medicine in metabolic health.
In-service heavy-haul railway bridges often operate under damaged conditions, requiring continuous health monitoring to ensure operational safety and facilitate intelligent maintenance. However, conventional assessment methods based on ground monitoring, limited by sparse sensor deployment, cannot fully capture the global bridge state and its evolution, and neglect stochastic track irregularities that induce significant response variability, thereby decreasing assessment reliability. To address these issues, a distribution-driven probabilistic bridge state assessment framework is proposed based on vehicle-bridge collaborative monitoring. The extreme-value distributions of vibration responses from the bogie, wheelset, and key span sections under stochastic track irregularities and varying bridge damage conditions are characterized using the probability density evolution method, revealing a mapping between damage-induced variations in bridge states and the corresponding distribution shifts. Based on this mapping, baseline thresholds for extreme responses are statistically determined under undamaged conditions to define six distinct bridge health levels. The probability for each level is estimated via a weight-adaptive hierarchical probabilistic evaluation model, which fuses probabilities derived from the extreme-value distributions of multi-source vibration response indicators, with weights allocated according to indicator sensitivity to damage. The bridge state is assessed as the one with the highest probability among the six levels. Case studies on scenarios with different damage locations and severities demonstrate that the proposed framework effectively distinguishes the effects of damage on bridge states and traces state evolution as damage progresses, providing reliable and interpretable assessments for bridge maintenance decision-making.
We present an end-to-end differentiable framework that facilitates three core capabilities in cardiovascular simulations: hemodynamic surrogate modeling, automated model calibration, and stochastic parameter tuning. Central to this approach is the introduction of a hybrid mechanistic and data-driven reduced order model (ROM) that represents each vascular domain through a nonlinear parametrization of lumped parameter networks. By exploiting the native differentiability of the pipeline, we calibrate the ROM parameters against a single high-fidelity 3D CFD simulation. The resulting optimized ROM serves as an efficient surrogate for both gradient-based deterministic and gradient-informed stochastic boundary condition calibration. With its computational efficiency and high fidelity, the framework directly addresses critical bottlenecks that currently limit the clinical adoption of cardiovascular simulations.
Several key questions remain unanswered regarding overparameterized learning models. It is unclear how (stochastic) gradient descent finds solutions that generalize well, and in particular the role of small random initializations. Matrix sensing, which is the problem of reconstructing a low-rank matrix from a few linear measurements, has become a standard prototypical setting to study these phenomena. Previous works have shown that matrix sensing can be solved by factorized gradient descent, provided the random initialization is extremely small. In this paper, we find that factorized gradient descent is highly robust to certain perturbations. This lets us use a perturbation term to capture both the effects of imperfect measurements, discretization by gradient descent, and other noise, resulting in a general formulation which we call perturbed gradient flow. We find that not only is this equivalent formulation easier to work with, but it leads to sharper sample and time complexities than previous work, handles moderately small initializations, and the results are naturally robust to perturbations such as noisy measurements or changing measurement matrices. Finally, we also analyse mini-batch stochastic gradient descent using the formulation, where we find improved sample complexity. This article is part of the theme issue 'Safe, secure and robust AI for safety-critical systems'.
Defect engineering via oxygen vacancy modulation has enabled a remarkable persistent photoconductivity effect in wide-bandgap semiconductors, spawning diverse artificial synapse architectures. However, stochastic defect distributions fundamentally limit device reproducibility and energy efficiency. Here, we advance defect engineering through an "ordered donor regulation" strategy, where Ga doping in single-crystalline ZnO microwires selectively passivates random oxygen vacancies, preserving nonvolatile memory while transforming transport from disordered defect-dominated to stable donor-regulated mode. This deterministic transition eliminates stochastic carrier-trapping kinetics, enabling precise conductance modulation at an ultralow bias of 300 μV; notably, a minimum energy consumption of 2.8 fJ per pulse is achieved at 1 mV─rivaling biological synapse efficiency. The device exhibits exceptional synaptic plasticity, characterized by a robust short-to-long-term memory transition; a nonvolatile retention time exceeding 15,000 s, as well as a 14% improvement in the paired-pulse facilitation index and an EPSC amplitude 3.3 times that of pristine ZnO. A three-layer neural network achieves 94.44% and 83.41% recognition accuracy on MNIST and Fashion-MNIST data sets, respectively. This work establishes ordered donor regulation as a paradigm for precision defect engineering in wide-bandgap semiconductor synapses, laying the material foundation for energy-efficient neuromorphic computing.
The allocation and timing for kidney matching and transplantation among incompatible donor-recipient pairs in the Paired Kidney Exchange face significant operational challenges of managing an uncertain waitlist. The time for dialysis of patients waiting for biologically compatible kidneys follows the concept of a stochastic queueing system. This study embodies the Learning Health System approach by integrating real-world data and mathematical modeling to generate actionable insights that guide decision-making. This work employs an M / M / 1 model with Poisson arrivals and exponentially distributed service times for incompatible donor-recipient pairs in a single transplant facility. Inter-arrival and service times are used to derive queue length and waiting-time measures to characterize system congestion and instability, with performance measures evaluated as indicative approximations analyzing the corresponding metrics. The system is persistently overloaded, with arrival rates exceeding service rates ( λ > μ ; ρ > 1), causing prolonged waiting times. Analytical and Hutson-B bootstrap confidence interval analysis shows congestion persists across plausible variations, and sensitivity analysis indicates moderate perturbations do not alleviate overload. Regression and correlation analyses indicate that delays are associated with arrival pressure and biologically imposed constraints, supporting compatibility-constrained matching as the primary operational challenge. Patients experience uncertain, extended waiting periods, reflecting stochastic instability in biologically constrained queueing systems. Transplant service management requires systematic reforms, as clinical matching and allocation delays hinder timely care. Establishing a national transplant registry, integrated real-time queue monitoring, decentralized planning with expansion of transplant centers and improved inter-center coordination for paired exchange and increasing its cycle lengths are essential to ensure efficient service delivery and informed healthcare policy. These interventions create a continuous feedback loop where operational data informs queueing-theoretic analysis, informs policy, system redesign, and ultimately improves patient access and outcomes, demonstrating how data-driven learning optimizes complex healthcare services.
Urban traffic congestion remains a critical challenge due to the inflexibility of conventional fixed-time or rule-based traffic signal control systems, which cannot adapt to real-time dynamic traffic flows. Existing deep reinforcement learning (DRL)-based adaptive signal control methods often neglect complex spatio-temporal dependencies in urban road networks, and graph-based spatio-temporal models (e.g., ASTGCN) are mostly limited to traffic flow prediction rather than signal control optimization. To address these gaps, this study proposes a novel adaptive traffic signal control framework by integrating Attention-based Spatio-Temporal Graph Convolutional Networks (ASTGCN) with Multi-Agent Deep Deterministic Policy Gradient (MADDPG). The ASTGCN module effectively captures spatial topological correlations and temporal dynamic patterns of traffic flow via graph attention and temporal convolution, while the MADDPG framework realizes centralized training and decentralized execution to solve multi-intersection cooperative control under partial observability. Experiments are conducted on the SUMO simulation platform with a four-intersection road network, using both real-world and stochastic traffic flows for validation. Comparative results with fixed-time control, Max-Pressure, centralized/distributed DRL, and vanilla MADDPG baselines show that the proposed method reduces traffic congestion by 20%-25% and achieves superior performance in convergence speed, generalization ability, extreme traffic load resilience, and non-uniform traffic adaptation. Ablation studies further verify the effectiveness of the ASTGCN-based spatio-temporal feature extraction component. This framework provides a scalable and data-driven solution for intelligent traffic signal control in urban traffic networks, supporting the development of smart mobility systems.