Traditional risk analysis methods in supply chain management have historically focused on the probability of adverse events occurring. While these approaches are useful for identifying potential threats, they often fall short in capturing the dynamic nature of resilience. Resilience is not merely about avoiding disruption; it is about the ability of a system to absorb impacts, adapt to changing conditions, and recover to a functional state. The research team, led by Shangbing Gao and colleagues from Huaian University and the Nanjing Hydraulic Research Institute, recognized that existing models did not adequately account for this adaptive capacity. Their work aims to bridge this methodological divide by developing a comprehensive evaluation framework that looks beyond simple probability to understand the full spectrum of supply chain performance under stress.

The proposed framework is structured around three fundamental dimensions of the fresh agricultural product supply chain: production, transportation, and consumption. By examining these three pillars, the model captures the entire lifecycle of the product, from the initial harvest to the final point of sale. This holistic approach ensures that no critical stage of the supply chain is overlooked. The study utilizes panel data from thirty-one provinces in China, spanning the period from 2009 to 2023. This extensive dataset provides a rich foundation for analyzing regional variations and temporal trends, allowing the researchers to observe how different areas respond to disruptions over time. The use of such a broad geographical scope is crucial for understanding the heterogeneity of supply chain resilience across diverse economic and climatic conditions.

At the core of the methodology is a hybrid modeling approach that integrates variable fuzzy set theory with machine learning. Variable fuzzy set theory is a mathematical tool that allows for the handling of uncertainty and imprecision in data. In the context of supply chains, where data can often be noisy or incomplete, this theory provides a robust way to assess resilience levels dynamically. The researchers employed this model to evaluate the current state of resilience, effectively categorizing different regions based on their ability to cope with disruptions. This assessment is not static; it is designed to reflect the changing conditions of the supply chain, providing a real-time snapshot of its health. The variable fuzzy recognition component is particularly valuable for distinguishing between different levels of resilience, offering a nuanced view that binary classifications cannot achieve.

Complementing the fuzzy set assessment is the use of a convolutional neural network, or CNN, for predictive modeling. While CNNs are traditionally associated with image processing, they have proven effective in analyzing sequential data, such as time series. In this study, the CNN is used to capture the temporal dynamics of the supply chain, identifying patterns and fluctuations that precede potential disruptions. The model is trained to recognize peak and off-peak variations in supply chain performance, allowing it to predict future resilience levels with high accuracy. This predictive capability is essential for proactive management, enabling stakeholders to anticipate problems before they escalate. The integration of CNNs with fuzzy set theory creates a powerful synergy, where the former provides predictive insight and the latter offers a robust assessment of current conditions.

Determining the relative importance of various indicators in the resilience assessment is a critical step in building a reliable model. The researchers employed a combined weighting method that integrates the Analytic Hierarchy Process, or AHP, with the CRITIC method. AHP is a structured technique for organizing and analyzing complex decisions, while CRITIC is an objective weighting method that considers the contrast intensity and correlation of indicators. By combining these two approaches, the study aims to balance subjective expert judgment with objective data-driven insights. This hybrid weighting scheme enhances the robustness and interpretability of the model, ensuring that the most relevant factors are given appropriate emphasis. The result is a more transparent and defensible assessment framework that can be trusted by policymakers and industry leaders.

The results of the study demonstrate the effectiveness of the proposed framework in distinguishing different resilience levels across regions. The model successfully identified areas with high resilience, which are better equipped to handle disruptions, and those with lower resilience, which are more vulnerable to shocks. This differentiation is crucial for targeted interventions, allowing resources to be allocated where they are most needed. Furthermore, the CNN model outperformed several baseline models, including Decision Tree, Random Forest, and Linear Regression, particularly in capturing nonlinear patterns and high-resilience cases. This superior performance highlights the advantage of using deep learning techniques for complex, dynamic systems like fresh agricultural supply chains. The model’s ability to handle noisy data and limited information further underscores its practical utility in real-world scenarios where perfect data is rarely available.

The implications of this research extend beyond academic interest, offering practical tools for early warning, risk mitigation, and decision-making in supply chain management. By providing a reliable means to assess and predict resilience, the framework empowers stakeholders to take proactive measures to protect the integrity of the food supply. This is particularly important in the face of increasing global uncertainties, from climate variability to geopolitical tensions. The study also highlights the importance of regional differences in resilience, suggesting that a one-size-fits-all approach to supply chain management is inadequate. Instead, tailored strategies that account for local conditions and vulnerabilities are necessary to ensure the stability of fresh agricultural product flows.

As the global food system continues to evolve, the integration of advanced computational methods into supply chain management will become increasingly vital. The work by Gao and colleagues represents a significant step forward in this direction, demonstrating the potential of combining fuzzy logic with deep learning to solve complex logistical challenges. The open access nature of the publication ensures that this knowledge is widely available, fostering further research and innovation in the field. By providing a robust and interpretable framework, the study contributes to a more resilient and sustainable food supply chain, ultimately benefiting consumers, producers, and the environment alike. The findings underscore the critical role of data-driven insights in navigating the complexities of modern agriculture and logistics.

Looking ahead, the application of such hybrid models could be expanded to other types of supply chains, including those for pharmaceuticals, electronics, and other sensitive goods. The principles of resilience assessment and prediction are universal, and the methodologies developed in this study can be adapted to various contexts. As machine learning techniques continue to advance, the potential for even more accurate and timely predictions grows. This research serves as a foundation for future work, inviting further exploration into the dynamics of supply chain resilience and the role of artificial intelligence in enhancing its management. In a world where the stability of food supplies is a matter of public health and economic security, such innovations are not just beneficial but essential.

Subject of Research: Resilience assessment and prediction of fresh agricultural product supply chains using variable fuzzy recognition and machine learning

Article Title: Resilience assessment and prediction of fresh agricultural product supply chains using variable fuzzy recognition and machine learning

Article References: Gao, S., Bai, A., Sun, X., Peng, A., Li, S., Zhao, W., & Ma, S. (2026). Resilience assessment and prediction of fresh agricultural product supply chains using variable fuzzy recognition and machine learning. Scientific Reports. https://doi.org/10.1038/s41598-026-75146-3

Image Credits: AI Generated

DOI: 10.1038/s41598-026-75146-3

Keywords: supply chain resilience, fresh agricultural products, machine learning, convolutional neural networks, fuzzy set theory, logistics, risk assessment, data analytics, food security, predictive modeling, Resilience, assessment

The system, described as a semantic-driven resilience decision engine, is built around a three-stage pipeline that mirrors how an experienced supply chain manager thinks. A perception layer reads raw text from news feeds, enterprise announcements and policy documents, and converts it into structured disruption vectors containing the event type, its intensity, the affected entities and the estimated duration. A decision layer takes those semantic signals and feeds them into a fuzzy-controlled version of the bat algorithm, a metaheuristic inspired by echolocating bats, which searches for the best combination of recovery actions. An execution layer then pushes the chosen strategy directly into the enterprise systems that run the factory floor: ERP for procurement and inventory, MES for production scheduling, WMS for warehousing, and TMS for transportation. Message queues decouple the layers asynchronously, so the whole loop can run continuously without waiting for each stage to finish.

The perception layer is built on a domain-adapted Chinese BERT model. Zhao pretrained it on 50,000 unlabeled domain texts collected over three years, then fine-tuned it on 10,500 labeled disruption records assembled from five manufacturing enterprises in automotive parts, home appliances and electronic information, plus 4,400 disruption-related news items gathered with keywords such as shutdown, fire, congestion and typhoon. The model uses a BERT-Span joint-extraction architecture in which a pointer network locates the start and end positions of entities such as suppliers, locations and materials, while relation classification identifies supply, location and causal links between them. On the test set, the module achieved an F1 score of 91.0 percent for entity extraction and 84.6 percent for relation extraction, with supplier entities performing best at 92.3 percent and causal relations hardest at 80.1 percent. Ablation experiments showed that domain-adaptive pretraining alone added 5.2 percentage points to entity extraction and 6.3 points to relation extraction, evidence that generic language models miss much of the specialized vocabulary of supply chain disruption.

The real innovation lies in how the extracted semantics drive the optimizer. In a standard bat algorithm, each candidate solution, visualized as a bat, updates its position and velocity based on a frequency parameter, while loudness and pulse rate control the balance between global exploration and local exploitation. These parameters are usually fixed or decay according to a preset schedule. Zhao’s fuzzy controller instead treats them as dynamic outputs of a 27-rule inference system with three inputs: the BERT-derived disruption intensity, the population diversity of the bat swarm, and the stagnation state, meaning the number of consecutive iterations without fitness improvement. Each input is mapped onto overlapping Low, Medium and High membership functions, and the activated rules are aggregated through Mamdani inference and centroid defuzzification into continuous adjustment factors that modify the loudness attenuation and pulse-rate enhancement coefficients every iteration.

Semantic information shapes the search at initialization as well as during iteration. The disruption intensity index, computed from four normalized components including event-type confidence, intensity-cue confidence, duration evidence and affected-scope evidence, controls the population size around a baseline of 40 bats. At a severe intensity score of 0.90, the population expands by roughly 36 percent, throwing more computational resources at the worst disruptions. The extracted disruption type also steers exploration: supplier events increase the mutation probability of alternative-supplier selection variables, while logistics events push the swarm toward transportation-route variables. This direct state-to-parameter mapping between external disruption semantics and internal search dynamics is what distinguishes the framework from earlier systems in which text analysis and optimization ran as separate, disconnected modules.

The optimizer evaluates candidate recovery strategies with a multi-objective fitness function that jointly weighs recovery time, additional cost and service-level maintenance, normalized to a common dimensionless scale. In the experiments the weights were set to 0.4 for recovery time, 0.3 for cost and 0.3 for service level, reflecting a recovery-speed-first but balanced objective, with a penalty term enforcing capacity, inventory and transportation constraints. The decision variables cover safety-stock adjustments, alternative-supplier selection and route replanning, and the search terminates when the global-best fitness shows no improvement for 20 consecutive generations. The author demonstrates that the retained best-fitness sequence is monotonically non-increasing and lower bounded, and therefore converges to a finite limit, while the bounded fuzzy control parameters prevent unbounded amplification during the search.

The performance gains are substantial. Across 30 independent runs, the fuzzy bat algorithm converged in 14.1 generations on average, 37.1 percent fewer than the standard bat algorithm’s 22.4, and achieved a mean fitness value 22.0 percent lower. Against newer swarm optimizers, the sparrow search algorithm and the dung beetle optimizer, the engine still reduced convergence generations by 16.6 percent, convergence time by 18.8 percent and fitness value by 6.4 percent relative to the dung beetle optimizer, with all Holm-corrected p-values below 0.001 and a Cohen’s d of 1.00 for fitness, a large effect size. Sensitivity analyses showed that fitness varied by less than 0.006 across population sizes from 30 to 50, and that the engine remained superior to manual and standard-BA baselines under every tested weighting of time, cost and service priorities.

End-to-end results under three realistic disruption scenarios were equally striking. Scenario A modeled a fire and production shutdown at a first-tier automotive parts supplier with an intensity score of 0.85; Scenario B modeled container congestion at a regional river port cutting turnover efficiency by 70 percent; and Scenario C combined a rainstorm closing a major expressway with a simultaneous secondary-supplier failure, producing the highest intensity score of 0.90. Compared with a manual decision-making baseline involving experienced supply chain managers, the engine reduced average recovery time by 37.2 percent, from 9.07 days to 5.70, cut additional cost by 36.1 percent, and raised the average service level by 10.5 percentage points to 90.9 percent. In the composite scenario, recovery time fell by 39.2 percent. An 18-case stress test spanning mild to severe intensities and normal to tight resource conditions confirmed that the advantage persists as disruptions worsen and recovery resources shrink, with performance degrading gradually rather than collapsing.

Engineering deployment data suggest the system is more than a simulation. Over a 72-hour continuous test on a hybrid GPU-CPU server, the engine sustained a steady-state text-processing throughput of 44 records per minute with a peak of 52, achieved an average end-to-end latency of 3.31 seconds with a 95th-percentile of 3.92 seconds, and completed 99.3 percent of 1,500 strategy calls successfully using retry and transaction-ID mechanisms. Latency decomposition showed the optimization stage dominates the response time at roughly 2.86 seconds, while semantic parsing takes about 298 milliseconds and execution-layer communication around 156 milliseconds. CPU utilization averaged 43.2 percent and memory stayed at 3.24 gigabytes with no leakage trend, indicating the model can run on standard industrial servers.

The study is candid about its limits. The perception module depends heavily on annotated domain data, dropping 5.2 F1 points without pretraining, and the current model is built on Chinese BERT, which may struggle with the multilingual stream of overseas supplier announcements and international logistics updates. All five validating enterprises sit in the same manufacturing region, and the disruption ontology covers supplier, logistics and weather events but not unforeseen cascading phenomena. The framework also optimizes for a single enterprise, even though one participant’s recovery measures can alter resource availability for everyone else. Future work outlined in the paper targets few-shot semantic extraction, multilingual models, open-set disruption detection, and multi-agent collaborative optimization combined with regional supply chain digital twins. For now, the results make a compelling case that when a language model’s reading comprehension directly tunes a swarm’s search behavior, factories can respond to chaos almost as fast as they read about it.

Subject of Research: An AI-driven supply chain resilience engine combining BERT event extraction with fuzzy bat optimization for manufacturing disruption response

Article Title: Semantic fuzzy bat optimization for regional manufacturing supply chain resilience

Article References: Zhao, Z. (2026). Semantic fuzzy bat optimization for regional manufacturing supply chain resilience. Discover Artificial Intelligence, 6(1), Article 1506. https://doi.org/10.1007/s44163-026-02488-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02488-2

Keywords: supply chain resilience, disruption management, BERT event extraction, fuzzy logic control, bat algorithm, swarm intelligence, regional manufacturing, natural language processing, multi-objective optimization, ERP integration, logistics, artificial intelligence

The review, conducted by Yuankai Zhang, Wei-Hua Lin, Qiong Tian and Xiangpei Hu, synthesizes nearly two decades of scholarship on business-to-consumer online order fulfillment. The authors systematically searched the literature covering work published from 2007 through May 2026. Their search initially surfaced 1,495 records, which they filtered down to 198 articles selected for detailed analysis. Rather than organizing the literature by academic subfield or methodology, the team anchored their classification to the real-world fulfillment process itself: order assignment, warehouse operations, shipping and delivery, and return operations. This process-oriented structure, the authors argue, mirrors the way decisions actually unfold inside an e-commerce company, making the review useful both as a scholarly reference and as a practical roadmap for operations managers.

The starting point of the chain is order assignment, the moment a retailer decides where each item in an order will be fulfilled. In modern e-commerce, this is rarely a trivial question. Large retailers operate networks with multiple fulfillment locations, including national distribution centers, regional warehouses and, increasingly under omni-channel models, physical stores that double as miniature fulfillment hubs. Assigning an order to the wrong node can mean longer delivery distances, higher shipping costs, split shipments that frustrate customers, or stock imbalances that leave popular items stranded in the wrong part of the network. The review shows that research on this stage has grown steadily more sophisticated as retailers’ networks have grown more distributed and as omni-channel strategies have blurred the traditional boundary between stores and warehouses.

Once an order lands at a fulfillment location, warehouse operations take over. This stage encompasses the physical heart of e-commerce: receiving inventory, storing it, picking the specific items a customer ordered, packing them and staging them for shipment. The review highlights why this stage is uniquely difficult in the online retail context. Unlike traditional retail, where a store sells a curated local assortment, B2C fulfillment must cope with very large product assortments, often spanning hundreds of thousands of stock-keeping units. Compounding the challenge, customer orders are typically personalized multi-item baskets rather than uniform bulk shipments, which means picking operations must handle enormous variety at high speed. The reviewed literature examines how these pressures shape decisions about warehouse layout, picking routing, batching of orders, zoning, and the deployment of automation, all of which interact with the upstream assignment decisions that determine what arrives at the warehouse floor in the first place.

The third stage, shipping and delivery, is where the retailer’s promises meet physical reality. Customers increasingly expect tight delivery time windows, sometimes measured in hours rather than days, and the review identifies these compressed windows as one of the four defining characteristics that make B2C fulfillment especially challenging. The other three are the very large product assortments already noted, the multi-location nature of fulfillment networks, and the personalized, multi-item composition of individual orders. Together, these four features distinguish online retail fulfillment from classical logistics problems and explain why the field has generated such a large and specialized body of research. Decisions about carrier selection, shipment consolidation, routing and delivery scheduling must all be made under these constraints, and each of them is constrained in turn by what happened, or failed to happen, in the warehouse hours earlier.

The final stage, return operations, is often treated as an afterthought in popular discussions of e-commerce, yet the review gives it full status as an integral part of the fulfillment process. Returns are a structural feature of online retail, not an anomaly: customers cannot physically inspect goods before buying, so a certain fraction of orders will always flow backward through the network. Handling those returns involves decisions about reverse logistics, inspection, restocking, refurbishing and disposition, and those decisions feed back into inventory positions and fulfillment capacity. The authors emphasize that the four stages they map are not silos but a connected system, and their review is structured to make those connections visible rather than burying them between disciplinary boundaries.

“B2C order fulfillment is not a collection of separate tasks. Decisions made at one stage can create consequences at another,” says Zhang. “By viewing the process as a connected system, we can better identify where coordination, data and new technologies can improve fulfillment decisions.” This systems perspective is the review’s central intellectual contribution. Much of the earlier literature, the authors find, examined individual stages in isolation: a paper might optimize picking routes without considering how order assignment shaped the pick list, or model delivery scheduling without accounting for warehouse cut-off times. The trend in recent research, and the direction the review charts for the future, is toward models that capture these interdependencies explicitly, coordinating decisions across the entire chain rather than optimizing each link separately.

That trend is closely tied to a second shift the review documents: the move from static planning toward integrated, real-time decision-making. Traditional fulfillment planning often assumed stable demand patterns and relatively long planning horizons, allowing decisions to be made in advance and executed as scheduled. Modern e-commerce operates at a tempo that makes such static assumptions increasingly untenable. Order streams arrive continuously, inventory positions change by the minute, and delivery promises are made dynamically at the moment of purchase. The reviewed literature shows researchers responding with approaches that update decisions in real time as new information arrives, embedding fulfillment logic directly into the customer-facing transaction rather than treating it as a back-office batch process.

Underpinning both shifts is the rapid maturation of data analytics and artificial intelligence. The team examined how data analytics, AI and emerging generative AI can support forecasting, scenario generation and operational decision-making across the fulfillment process. Demand forecasting, long a staple of operations research, is being augmented by machine learning models that can exploit far richer behavioral data, which the review identifies as a promising direction under the heading of behavioral analytics. Generative AI, meanwhile, opens possibilities for scenario generation, allowing planners to stress-test fulfillment networks against synthetic but plausible demand patterns. Importantly, the authors frame these technologies not as replacements for classical operations research. “Rather than replacing optimization methods, these technologies are increasingly viewed as decision-support tools that can work alongside optimization,” Zhang adds. In this hybrid vision, machine learning supplies predictions and scenarios, while optimization translates them into concrete, constraint-respecting operational plans.

The review also points toward several forward-looking research frontiers beyond real-time integration and AI. Decision-making under uncertainty remains a persistent challenge, since fulfillment networks must cope with demand spikes, supply disruptions and weather-related delivery delays that no forecast fully anticipates. Sustainability is flagged as a growing concern, encompassing the emissions footprint of expedited shipping, packaging waste and the energy profile of large-scale warehouse automation. And the authors highlight the potential of combining artificial intelligence with digital twins and optimization: virtual replicas of fulfillment networks that can be continuously updated with real operational data and used to test decisions before they are executed in the physical world. For managers, the practical message of the review is to stop evaluating warehouse, delivery and returns decisions in isolation and instead instrument and optimize the entire chain. For researchers, it is a call to build models that match the connected, data-saturated, time-critical reality of modern online retail. For the millions of shoppers who will click “Place Order” today without a second thought, it is a reminder that behind that single click lies a cascade of interlocking decisions, and a research community working to make that cascade faster, cheaper and more responsive.

Subject of Research: B2C e-commerce order fulfillment decision-making across order assignment, warehouse operations, delivery and returns

Article Title: From click to doorstep: the hidden decisions behind e-commerce orders

Article References: From click to doorstep: the hidden decisions behind e-commerce orders. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: e-commerce, order fulfillment, supply chain, warehouse operations, logistics, omni-channel retail, artificial intelligence, optimization, digital twins, returns management, systematic review, operations research

Researchers Yanxia Ma and Yufeng Zou of Shanghai Sipo Polytechnic have developed a deep learning pipeline that identifies abnormal worker behaviors in warehouse surveillance footage, reporting their results in the open-access journal Discover Artificial Intelligence. Their system chains together three neural network components: one that finds and locates people in each video frame, one that converts each detected person into a stick-figure skeleton of seventeen joints, and one that reads the motion of that skeleton over consecutive frames to decide whether the person is operating normally, falling, colliding with a colleague, or climbing on storage racks.

The first stage relies on YOLOv5, a member of the You Only Look Once family of object detectors. The authors deliberately chose version 5 over its newer successors, YOLOv6 through YOLOv8, arguing that the later versions’ structural refinements buy extra accuracy at the cost of substantially longer training times, which slows down model iteration and deployment. For warehouse monitoring, where small targets and rapid retraining matter more than squeezing out the last percentage point of precision, YOLOv5 struck the right balance. Architecturally, the model splits into a backbone that extracts convolutional features at multiple scales using CBS modules, C3 bottleneck blocks and a spatial pyramid pooling fast module; a neck that fuses features from different scales through upsampling; and a head that regresses candidate bounding boxes back onto the original image coordinates to localize each worker.

Once a worker is boxed, the second stage extracts their pose using a two-branch, multi-stage network in the style of OpenPose. One branch predicts part confidence maps, heat maps that estimate where each skeleton joint sits in the image, while the other predicts part affinity fields, vector maps that encode the direction in which adjacent joints should be connected. Each stage refines the estimates of the previous one, with the original image features fed in repeatedly, and training is driven by a summed loss that compares predicted heat maps and affinity fields against manually annotated ground truth, masked by a binary function that ignores unannotated regions. YOLOv5 supplies the image features inside each bounding box, tying the two stages together.

The third stage classifies action types from skeletal data accumulated over consecutive frames. The researchers normalize joint coordinates, remove head joints and fill in missing points, then derive features including joint positions and movement speeds. These flow into a back-propagation neural network, a classic fully connected architecture with input, hidden and output layers whose hidden-layer activation functions fit the nonlinear patterns linking skeleton dynamics to behavior labels.

To test the system, the team drew on a week of surveillance video from a logistics company recorded in November 2025, covering warehousing zones, goods handling areas and shelf areas, with footage of normal operations as well as genuine falls, collisions and illegal climbing. After discarding blurred, occluded or empty segments and sampling one frame in every thirty, they assembled 20,000 valid images, split into 12,000 training and 8,000 test samples, with single-person and multi-person scenarios kept separate throughout. Annotators marked bounding boxes, seventeen key joints from head to ankles with connecting limb lines, and behavior labels; a cross-review pass removed mislabeled samples. Experiments ran on a laboratory server with a Core i7 processor, an RTX 4060 graphics card and 32 gigabytes of memory.

The benchmarks pitted each stage of the proposed system against established alternatives. For person detection and localization, YOLOv5 outperformed R-CNN, YOLOv3 and a fully convolutional neural network, and it was notably the most stable when moving from single-person to multi-person images, a scenario in which the competitors’ performance visibly declined. For skeleton extraction, the proposed pose network produced smaller joint-point errors than the CPM and CPN algorithms in both scenarios. For action classification, the BPNN beat both a support vector machine and a random forest, with the SVM faring worst. Performance was quantified with precision, recall, F-score and intersection-over-union, with each algorithm tested three times and results reported as means with standard deviations.

Ablation experiments underscored why the full pipeline matters. Stripping out YOLOv5 and asking the skeleton network to work on raw images degraded results, because cluttered warehouse backgrounds and passing colleagues inject noise that accurate framing of the target removes. Removing skeleton extraction and classifying from raw image features hurt even more, confirming that articulated pose is the core feature representation. And swapping the BPNN for an SVM under identical front-end modules lowered classification quality, which the authors attribute to the neural network’s stronger nonlinear fitting on high-dimensional skeleton time series, particularly in crowded scenes.

The practical payoff, the researchers argue, is a shift from reactive to proactive warehouse safety. A fall can leave an injured worker unattended while blocking colleagues; collisions damage goods and sour teamwork; climbing racks loads shelving beyond its design and can trigger collapses and property losses. A system that flags these behaviors the moment they appear on camera can summon help to a fallen worker, mediate a collision before it escalates, and warn off anyone scaling the shelves, all without waiting for a human supervisor to happen to be watching the right monitor at the right moment.

Limitations remain. The validation rested on one week of footage from a single company processed in simulation on modest hardware, and the study focused exclusively on personnel behavior rather than the fire, shelf-structure and dangerous-goods scenarios the authors identify as the other pillars of intelligent warehouse safety. Still, by fusing detection, pose estimation and classification into one stable pipeline that holds up in multi-person images, the work sketches a plausible near future in which the same cameras that watch inventory also watch out for the people moving it.

Subject of Research: Deep learning-based recognition of abnormal personnel behaviors in e-commerce warehouse safety management

Article Title: Research on intelligent safety management of e-commerce logistics storage through a deep learning model

Article References: Ma, Y., & Zou, Y. (2026). Research on intelligent safety management of e-commerce logistics storage through a deep learning model. Discover Artificial Intelligence, 6(1), Article 1347. https://doi.org/10.1007/s44163-026-02236-6

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02236-6

Keywords: deep learning, YOLOv5, skeleton extraction, action recognition, warehouse safety, e-commerce logistics, surveillance, computer vision, BPNN, pose estimation, abnormal behavior detection, logistics

The researchers’ starting point is the ant colony system, a family of algorithms inspired by the way real ants find short paths to food. Individual ants deposit pheromone along the trails they walk, and subsequent ants preferentially follow stronger trails, reinforcing good routes while weaker ones evaporate. Translated into computation, artificial ants construct delivery routes step by step, guided by two signals: pheromone trails that encode collective experience about which customer-to-customer moves have historically produced good routes, and heuristic information that estimates the immediate desirability of each possible move. Over many iterations, the colony converges toward high-quality solutions without ever enumerating the full search space. Since Marco Dorigo introduced ant colony optimization in the early 1990s, the approach has been applied to everything from network routing to scheduling, and it remains a workhorse for routing problems where exact methods run out of steam.

What distinguishes the new algorithm is the way it deploys not one but four heuristic techniques inside the colony’s decision-making machinery. In a conventional ant colony system, each ant typically relies on a single heuristic rule when choosing the next customer to visit. That rigidity can be a liability, because different heuristics excel in different regions of the search: one might be better at deciding which customer to add to a route, another at deciding where to insert a customer into an existing route, and still others at repairing or restructuring routes that have become inefficient. Rather than betting on a single rule, the authors let the algorithm learn which heuristic serves it best at any given moment, drawing on the complementary strengths of all four.

The mechanism that orchestrates this cooperation is a ranking-based selection procedure built on the classic roulette wheel. Each of the four heuristics is assigned a weight reflecting how well it has performed recently, and when the algorithm needs to make a decision, it spins the metaphorical wheel, giving better-ranked heuristics proportionally larger slices. Crucially, the weights are not fixed. As the search progresses, heuristics that consistently contribute to improved solutions gain a larger share of future decisions, while those that fail to help are gradually demoted. This adaptive allocation means the algorithm effectively conducts a continuous, built-in experiment, reallocating its computational effort toward whichever strategy is currently paying off. The result, according to the authors, is the generation of promising solutions that no single heuristic could reliably produce alone.

A second design element addresses one of the most persistent failure modes of swarm intelligence: stagnation. Because pheromone reinforcement favors routes that have already been found good, a colony can settle into a rut, repeatedly exploring the same small set of routes while better alternatives elsewhere in the search space go unnoticed. The new algorithm counters this by incorporating a no-improvement counter into its solution selection process. When the counter indicates that the colony has gone a stretch of iterations without finding anything better, the algorithm adjusts its selection behavior to push the search toward less-explored territory, restoring diversity among the candidate solutions. This balance between intensification, which exploits known good regions, and diversification, which explores new ones, is the central tension of all metaheuristic design, and the counter gives the algorithm an explicit, self-monitoring handle on it.

The third and perhaps most inventive contribution is a new local improvement operator the authors call 2-opt-special, inspired by the venerable 2-opt move that has been a staple of routing heuristics since the 1960s. In its classic form, 2-opt takes a route, removes two edges, and reconnects the remaining segments in the alternative way, undoing any crossings and shortening the tour. The special variant introduced here adapts and extends this idea to bolster both the efficiency and the effectiveness of the ant colony framework, giving the colony a targeted tool for refining the routes its ants construct. Local search operators of this kind matter enormously in practice: constructive methods such as ant colonies are good at assembling plausible routes, but the difference between a plausible route and an excellent one often lies in small, surgical repairs that only a well-designed neighborhood move can deliver.

To find out whether all this machinery actually works, the team ran extensive computational experiments on well-known benchmark instances containing 100 and 400 customers, the standard proving grounds for VRPTW algorithms. They compared their method against other ant colony-based algorithms and against established solutions from the literature. The results, the authors report, affirm that the proposed algorithm is effective and competitive in addressing VRPTW challenges, holding its own against prior approaches across the benchmark suite. The comparisons were supported by the kind of statistical methodology that has become standard practice for evaluating evolutionary and swarm algorithms, following nonparametric testing procedures designed to compare stochastic optimizers fairly rather than relying on single lucky runs.

The significance of the work extends beyond the leaderboard. Vehicle routing with time windows sits at the heart of logistics, e-commerce, and emergency transportation, and even percentage-point improvements in route quality translate into real savings in fuel, fleet size, and emissions. The paper’s emphasis on combining multiple heuristics under an adaptive selection scheme reflects a broader trend in metaheuristics research: rather than designing ever more exotic single algorithms, researchers increasingly build frameworks that can assemble and tune proven components on the fly. The authors’ own research trajectory points in the same direction, including earlier work on hybrid adaptive large neighborhood search for capacitated routing problems and on metaheuristic techniques applied to engineering design problems.

For the field, the study also offers a reminder of how much headroom remains in a technique that is more than three decades old. Ant colony optimization was conceived as a simple model of distributed learning, yet its modern descendants now incorporate ranking mechanisms, adaptive strategy portfolios, stagnation detection, and specialized local search, layering decades of algorithmic insight onto the original biological metaphor. The version of the algorithm described here, published in volume 38 of Neural Computing and Applications as article number 705, was received in October 2024 and accepted in July 2026, and the authors declare no conflicts of interest and no specific funding for the work.

As delivery networks grow denser and customer expectations for precise delivery windows tighten, the computational problems underlying logistics will only intensify. Approaches like the one developed by Kuyu, Fajardo-Calderin, and Onieva suggest that the path forward lies not in any single clever rule but in flexible systems that know when to switch tactics, when to double down on what works, and when to break out of a rut. Virtual ants, it turns out, still have plenty to teach us about finding the best way home.

Subject of Research: A novel ant colony system using multiple heuristic strategies for solving the vehicle routing problem with time windows

Article Title: A novel ant colony system for vehicle routing problem with time windows: leveraging multiple heuristic strategies for improved solutions

Article References: Kuyu, Y. Ç., Fajardo-Calderin, J., & Onieva, E. (2026). A novel ant colony system for vehicle routing problem with time windows: leveraging multiple heuristic strategies for improved solutions. Neural Computing and Applications, 38(17), Article 705. https://doi.org/10.1007/s00521-026-12418-z

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12418-z

Keywords: ant colony optimization, vehicle routing problem, time windows, heuristics, metaheuristics, swarm intelligence, combinatorial optimization, logistics, local search, 2-opt, benchmark instances, Neural Computing and Applications

The classical transportation problem, first formulated by F. L. Hitchcock in 1941, asks a deceptively simple question: what is the cheapest way to ship a product from several sources to several destinations while respecting supply and demand limits? For decades, extensions of this model have underpinned routing decisions in manufacturing, retail, and humanitarian relief. But the classical formulation demands precise numbers: exact unit costs, exact capacities, exact demands. In reality, fuel prices fluctuate, weather and congestion stretch transit times, demand forecasts are educated guesses, and route safety is a matter of expert judgment rather than a measured constant. Deterministic models fed with point estimates produce plans that are optimal only for one idealized version of the world and can become infeasible or unexpectedly costly when reality deviates.

The new framework departs from that tradition by representing every imprecise quantity, including costs, capacities, supplies, demands, transit times, and security levels, as trapezoidal fuzzy numbers. Each such number is defined by four values: a lower support marking the smallest plausible value, a core interval of most-likely values, and an upper support marking the largest plausible value. Instead of forcing an expert to name a single figure for the cost of shipping a pallet by rail, the model accepts a range with graded confidence, mirroring how human judgment actually works. The authors argue this is a deliberate choice: logistics quantities are imprecise but not necessarily random, so they are naturally described by possibility distributions rather than the probability distributions that stochastic programming presupposes. Compared with robust optimization, which protects against a worst case defined by an uncertainty set that is itself hard to specify, the fuzzy formulation occupies a pragmatic middle ground.

Three features distinguish the model from prior fuzzy transportation research, and the authors present them as a package whose combined properties had not previously been established. First, it retains the full solid, three-dimensional structure of the problem, meaning decisions span sources, destinations, and conveyance modes simultaneously, so multi-modal trade-offs between road, rail, air, and sea are captured directly rather than aggregated away. A mode-exclusivity constraint ensures that at most one conveyance mode serves any origin-destination pair, eliminating a semantic ambiguity in earlier formulations in which several modes could implicitly serve the same link and over-count delivery times. Second, it optimizes three conflicting objectives at once: minimizing total cost, minimizing total transit time, and maximizing route security, where security is defined as the reliability of delivering cargo intact and on schedule, quantified on a normalized scale from zero to one and aggregated from incident rates, infrastructure quality, and expert assessment. To the authors’ knowledge, no prior fuzzy solid transportation model treats route reliability as a first-class optimization target rather than a hard constraint or an afterthought.

Third, and crucially for computation, the framework converts fuzzy data into crisp equivalents using centroid defuzzification, which for a trapezoidal number simply averages its four defining values. The choice is not arbitrary. Centroid defuzzification is a linear operator: the expected value of a sum equals the sum of expected values, and scaling a fuzzy number scales its centroid proportionally. Substituting expected values therefore keeps every objective and constraint linear, preserving the model as a mixed-integer linear program solvable to global optimality with standard solvers. Non-linear or rank-dependent defuzzification methods would destroy this linearity and, with it, the convexity guarantees the authors establish. For symmetric fuzzy data, the centroid coincides with the signed-distance and graded-mean methods, making the optimal plan invariant to the choice among them; for asymmetric data, preliminary checks showed the dominant flow structure of the Pareto front remains unchanged.

To handle the three competing objectives, the study employs the weighted Tchebycheff scalarization method. Each objective is normalized between its ideal value, obtained by optimizing it alone, and its nadir value, estimated from a payoff table. The scalarized problem then minimizes the largest weighted deviation from the ideal points. Unlike simple weighted-sum approaches, the Tchebycheff method can recover Pareto-optimal solutions lying on non-convex regions of the trade-off surface, and by systematically sweeping the weights over the unit simplex, the method traces out a well-distributed Pareto front. The authors go beyond numerics to prove analytical properties: the constraint polytope is non-empty after balancing, the feasible region is convex for any fixed mode selection, the expected-value operator is Lipschitz-continuous in the fuzzy parameters so small data perturbations produce proportionally bounded changes in the optimum, and with strictly positive weights every solution of the scalarized sub-problem is Pareto optimal, with the complete front recoverable as the weight grid is refined.

Two hypothetical numerical examples demonstrate the framework in action. The first models a manufacturing supply chain with three suppliers, three distribution centers, and two modes, road and rail. After defuzzification revealed total supply exceeding total demand, a dummy destination with zero cost restored balance. The resulting plan cost roughly $1,250 less than a plan built under deterministic assumptions, kept transit times within acceptable limits, and maintained high security across all selected routes. Benchmarking against a classical deterministic cost-minimizing program, a single-objective fuzzy cost model, and a two-objective fuzzy goal program proved revealing: the deterministic model achieved the lowest nominal cost, but when unit costs were increased by fifteen percent its plan became the most expensive of the four. The proposed model accepted a cost premium of about five percent in exchange for explicit security optimization and demonstrably more robust plans.

The second example turns to disaster-relief logistics, a domain where uncertainty is not a nuisance but the defining condition. Two supply depots serve three affected districts using air, land, and sea transport, with all parameters modeled as fuzzy numbers to reflect the chaos of crisis environments. The optimal plan, costing about $3,400 with transit times around six hours and security levels above 0.85, prioritized air transport for the most urgent and insecure region while assigning cheaper sea and land modes to less critical destinations. The authors emphasize that decision-makers can adjust the objective weights, for instance raising security and time weights during a disaster, to obtain faster and safer plans at a quantified cost premium, with the model producing revised plans that still respect the fuzzy constraints.

Sensitivity analysis reinforced the theoretical claims. When the fuzzy cost parameters were perturbed systematically, the optimal cost responded almost linearly, with no abrupt jumps, confirming the Lipschitz-stability result numerically. Sweeping the cost weight traced the cost-time-security trade-off continuously, and a ten percent demand growth required only reallocation rather than wholesale re-planning, indicating that plans degrade gracefully. Computationally, the approach is practical: for instances up to three sources, three destinations, and three modes, each scalarized linear program solved in well under a second, and a full Pareto front over a twenty-one-point weight grid was generated in a few seconds. A larger synthetic instance with 120 route triples still terminated in seconds using branch-and-cut, and because the continuous relaxation retains the network structure of the classical transportation problem, the authors expect the formulation to scale gracefully to hundreds of nodes.

The study is not without acknowledged limitations: defuzzification inevitably discards some information, and genuinely large-scale benchmark instances and hybrid fuzzy-stochastic extensions remain future work. Yet the contribution stands as a template for decision-making under ambiguity. Rather than issuing a single prescriptive plan, the model hands managers a menu of Pareto-optimal options, each defensible in the trade-off it makes between money, speed, and safety. For a relief agency weighing lives against budgets, or a manufacturer hedging against volatile fuel prices, that transparency may prove as valuable as the mathematics behind it.

Subject of Research: Fuzzy multi-objective optimization of multi-modal transportation and logistics planning under uncertain cost, time, and security parameters

Article Title: Fuzzy multi objective solid transportation model for optimal distribution

Article References: Suryawanshi, G., Muley, A., & Fegade, M. (2026). Fuzzy multi objective solid transportation model for optimal distribution. Discover Informatics, 1(1), Article 24. https://doi.org/10.1007/s44564-026-00026-x

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00026-x

Keywords: fuzzy optimization, solid transportation problem, multi-objective optimization, supply chain management, logistics, trapezoidal fuzzy numbers, defuzzification, Pareto front, disaster relief logistics, route security, weighted Tchebycheff method, linear programming

The researchers call their framework ADAPT-FUSE, short for Adaptive Deep Attention Prediction Transformer with Fuzzy Unified Stacking Ensemble. The name is a mouthful, but the design philosophy is straightforward: no single machine learning paradigm captures every kind of structure in supply-chain data, so why not combine several? The five-layer architecture runs five parallel feature extractors over each transaction. A temporal convolutional network uses dilated causal convolutions to pick up local, translation-invariant temporal motifs at low computational cost. A bidirectional LSTM encoder reads the sequence forwards and backwards through gated recurrent memory, capturing order-sensitive dependencies that convolution cannot model. A multi-head self-attention transformer computes scaled dot-product attention across all feature positions simultaneously, exposing global pairwise interactions regardless of distance.

The remaining two branches address structure and uncertainty. A graph neural network represents the supply chain as a directed graph whose nodes are origin provinces, platforms, logistics providers and destination countries, with edges weighted by historical transaction volumes; message-passing aggregates neighbourhood information so that carrier reliability and network topology inform each prediction. Finally, a fuzzy inference system maps crisp inputs into triangular membership functions labelled Low, Medium and High, applies Mamdani-style rule aggregation and defuzzifies the result by the centroid method, injecting graded, rule-based uncertainty that the differentiable modules do not express explicitly. The five outputs are projected to a common dimension and combined through an adaptive gated fusion mechanism, in which a softmax-normalised gate scores each modality per transaction, allowing the model to up-weight or down-weight each view depending on the input rather than relying on fixed weights.

The fused representation then feeds a stacking ensemble. Four base learners — LightGBM, XGBoost, ExtraTrees and a compact multilayer perceptron — produce out-of-fold probability predictions, which are weighted by a fuzzy confidence function that rewards predictions far from the 0.5 decision boundary and discounts uncertain ones. A gradient-boosting meta-learner trained on these confidence-weighted meta-features produces the final classification, calibrated via Platt scaling. Hyperparameters of both base learners and the meta-learner are tuned by a dual optimisation framework combining Bayesian optimisation with Tree-structured Parzen Estimation and a multi-objective genetic algorithm. The authors argue this is the first such integrated effort in the cross-border e-commerce logistics domain, and the modular design means each extractor can in principle be swapped or retrained independently.

Crucially, the evaluation rests on a fully synthetic benchmark rather than proprietary platform data. The dataset, named C-CBEC-SC, contains 5,000 machine-generated transactions spanning eight major Chinese export provinces — Guangdong, Zhejiang, Fujian, Shanghai, Jiangsu, Shandong, Beijing and Chongqing — over 2021 to 2024, with 24 raw features and 5 engineered composites. Marginal distributions were calibrated only to publicly available aggregate trade and logistics statistics. The label-generation rule is documented in full: a latent utility built from standardised transit, customs and warehouse times, forecast accuracy, entity quality effects and technology-adoption terms, including deliberate interaction effects such as joint AI–IoT adoption and the amplification of disruption on long routes. Substantial Gaussian noise caps attainable accuracy well below perfection, which is why results hover near 0.89. The authors are explicit that any feature-importance finding describes the generator, not real Chinese e-commerce operations.

On the held-out test partition, ADAPT-FUSE achieved the best accuracy among eight evaluated models at 0.8940, along with the highest macro recall of 0.8504 and macro F1-score of 0.8543, with macro precision of 0.8585 and an AUC of 0.9441 essentially level with the strongest baseline. Five-fold stratified cross-validation yielded a more conservative mean accuracy of 0.8752 plus or minus 0.0095, which the authors regard as the more honest estimate. The recall advantage is the operationally meaningful one: in delay-risk alerting, the minority class of delayed shipments is the class worth catching, since an unflagged delay produces refund exposure and customer churn while a false alarm costs only a review step. A model that improves minority-class recall at comparable ranking quality is therefore preferable even when overall accuracy differences are small.

What distinguishes the paper is its statistical candour. A McNemar test comparing ADAPT-FUSE against the strongest baseline, Deep-LSTMNet, produced a p-value of 0.5044, and bootstrap confidence intervals for the AUC and F1 differences both spanned zero. The authors state plainly that the margin over the strongest baseline is not statistically significant, that the 1,000-record test set is underpowered to resolve differences of this magnitude, and that they have removed any claim of significant outperformance from their conclusions. They frame the contribution as a competitive, fully reproducible architecture rather than a performance breakthrough — a transparent, checkable negative-to-neutral significance result that they argue is more useful to the field than an unverifiable positive one.

The ablation study adds further nuance. Removing the fuzzy inference and recurrent branches measurably degrades performance, marking them as the workhorses of the architecture. But removing the transformer attention block leaves accuracy and F1 essentially unchanged, and removing the graph neural network module changes accuracy by less than 0.002, because the relational information it encodes is partly recoverable from the one-hot entity columns available to every model. Rather than insisting all five paradigms are indispensable, the authors conclude that component value is data-dependent: the attention and graph branches may matter far more on larger datasets with richer topology, but on 5,000 synthetic records their marginal contribution is small. The model trains in 10.5 seconds on 4,000 records and serves predictions at roughly 0.014 milliseconds per record, comfortably inside real-time logistics budgets.

The authors also acknowledge that the graph module uses a static topology fixed at training time. Under sudden structural change — a port closure, a carrier suspension, a rerouted corridor — the entity encodings become stale, and errors would concentrate on the affected corridor until retraining. They propose operational monitoring triggers that widen predictive intervals and route affected shipments to human review, and methodological upgrades such as temporal graph networks with time-varying adjacency. A three-stage deployment pathway is outlined: standardised near-real-time data integration with strict exclusion of two outcome-adjacent composite features to prevent leakage, containerised model serving validated by A/B comparison, and continuous retraining on rolling windows with drift monitoring. Notably, the team withdrew an earlier projection of a 3 to 7 percentage-point on-time-rate gain, conceding that no such figure can be responsibly estimated without a live trial on real operations.

The study’s limitations are stated with equal precision: the data are synthetic, the sample is modest, two of five branches contribute little at this scale, the graph topology is untested under disruption, interpretability is only partial, and the baselines are faithful re-implementations of algorithmic families rather than original author code. Yet the release of the generator, model code, seeds, baseline configurations and raw result logs means every reported number can be reproduced line by line. Future work targets validation on large real operational datasets, temporal graph learning, multi-modal inputs such as satellite imagery and real-time weather feeds, federated training across platforms without data sharing, and per-prediction explanation modules extending the fuzzy-rule traces. In a field often criticised for overclaiming, ADAPT-FUSE may be remembered less for its metrics than for demonstrating how an AI paper can be ambitious, rigorous and honest at the same time.

Subject of Research: Hybrid deep learning architecture for on-time delivery prediction in cross-border e-commerce supply chains

Article Title: Computing intelligent models for cross-border E-commerce supply chain optimization using artificial intelligence

Article References: Jiao, Z., & Zhang, X. (2026). Computing intelligent models for cross-border E-commerce supply chain optimization using artificial intelligence. Discover Artificial Intelligence, 6(1), Article 1318. https://doi.org/10.1007/s44163-026-02292-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02292-y

Keywords: cross-border e-commerce, supply chain optimization, artificial intelligence, deep learning, ADAPT-FUSE, graph neural networks, fuzzy inference, stacking ensemble, transformer attention, delivery prediction, logistics, synthetic benchmark