Low Computational Resources Research Articles

Atherosclerotic plaque classification in carotid ultrasound images using machine learning and explainable deep learning

ObjectiveThe incidence of cardiovascular diseases (CVD) is rising rapidly worldwide. Some forms of CVD, such as stroke and heart attack, are more common among patients with certain conditions. Atherosclerosis development is a major factor underlying cardiovascular events, such as heart attack and stroke, and its early detection may prevent such events. Ultrasound imaging of carotid arteries is a useful method for diagnosis of atherosclerotic plaques; however, an automated method to classify atherosclerotic plaques for evaluation of early-stage CVD is needed. Here, we propose an automated method for classification of high-risk atherosclerotic plaque ultrasound images. MethodsFive deep learning (DL) models (VGG16, ResNet-50, GoogLeNet, XceptionNet, and SqueezeNet) were used for automated classification and the results compared with those of a machine learning (ML)-based technique, involving extraction of 23 texture features from ultrasound images and classification using a Support Vector Machine classifier. To enhance model interpretability, output gradient-weighted convolutional activation maps (GradCAMs) were generated and overlayed on original images. ResultsA series of indices, including accuracy, sensitivity, specificity, F1-score, Cohen-kappa index, and area under the curve values, were calculated to evaluate model performance. GradCAM output images allowed visualization of the most significant ultrasound image regions. The GoogLeNet model yielded the highest accuracy (98.20%). ConclusionML models may be also suitable for applications requiring low computational resource. Further, DL models could be more completely automated than ML models.

Head-Free Lightweight Semantic Segmentation with Linear Transformer

Existing semantic segmentation works have been mainly focused on designing effective decoders; however, the computational load introduced by the overall structure has long been ignored, which hinders their applications on resource-constrained hardwares. In this paper, we propose a head-free lightweight architecture specifically for semantic segmentation, named Adaptive Frequency Transformer (AFFormer). AFFormer adopts a parallel architecture to leverage prototype representations as specific learnable local descriptions which replaces the decoder and preserves the rich image semantics on high-resolution features. Although removing the decoder compresses most of the computation, the accuracy of the parallel structure is still limited by low computational resources. Therefore, we employ heterogeneous operators (CNN and vision Transformer) for pixel embedding and prototype representations to further save computational costs. Moreover, it is very difficult to linearize the complexity of the vision Transformer from the perspective of spatial domain. Due to the fact that semantic segmentation is very sensitive to frequency information, we construct a lightweight prototype learning block with adaptive frequency filter of complexity O(n) to replace standard self attention with O(n^2). Extensive experiments on widely adopted datasets demonstrate that AFFormer achieves superior accuracy while retaining only 3M parameters. On the ADE20K dataset, AFFormer achieves 41.8 mIoU and 4.6 GFLOPs, which is 4.4 mIoU higher than Segformer, with 45% less GFLOPs. On the Cityscapes dataset, AFFormer achieves 78.7 mIoU and 34.4 GFLOPs, which is 2.5 mIoU higher than Segformer with 72.5% less GFLOPs. Code is available at https://github.com/dongbo811/AFFormer.

Studies advanced in chatbots based on deep learning

Chatbots have always been a hot research topic in the field of human-computer interaction research, which aims to build a conversational intelligent response model to simulate human dialogue. Thanks to the rapid development of natural language processing technology and the continuous accumulation of dialogue data, the research of chat robots have made remarkable progress, which has gradually been widely used in various fields such as e-commerce and smart home. According to different technical frameworks, existing chatbots are mainly divided into two types: retrieval chatbots and generative chatbots. As the primary means of implementing chatbots in the industry, retrieval chatbots have smooth responses and low computational resource consumption. In contrast, generative chatbots do not require a predefined knowledge base and can dynamically generate responses based on the dialogue content. In this paper, focusing on the above two types of frameworks, we introduce the latest research progress in the field of deep learning-based chatbots in detail, including the representative algorithms and corresponding pipelines. Second, we compare the performance of representative algorithms on different datasets. We also summarize the problems chatbot technology research faces and give an outlook on its future development trends.

Low-Resource Generation Method for Few-Shot Dolphin Whistle Signal Based on Generative Adversarial Network

Dolphin signals are effective carriers for underwater covert detection and communication. However, the environmental and cost constraints terribly limit the amount of data available in dolphin signal datasets are often limited. Meanwhile, due to the low computational power and resource sensitivity of Unmanned Underwater Vehicles (UUVs), current methods for real-time generation of dolphin signals with favorable results are still subject to several challenges. To this end, a Masked AutoEncoder Generative Adversarial Network (MAE-GAN) model is hereby proposed. First, considering the few-shot condition, the dataset is extended by using data augmentation techniques. Then, to meet the low arithmetic constraint, a denoising autoencoder with a mask is used to obtain latent codes through self-supervised learning. These latent codes are then utilized in Conditional Wasserstein Generative Adversarial Network-Gradient Penalty (CWGAN-GP) to generate a whistle signal model for the target dataset, fully demonstrating the effectiveness of the proposed method for enhancing dolphin signal generation in data-limited scenarios. The whistle signals generated by the MAE-GAN and baseline models are compared with actual dolphin signals, and the findings indicate that the proposed approach achieves a discriminative score of 0.074, which is 28.8% higher than that of the current state-of-the-art techniques. Furthermore, it requires only 30.2% of the computational resources of the baseline model. Overall, this paper presents a novel approach to generating high-quality dolphin signals in data-limited situations, which can also be deployed on low-resource devices. The proposed MAE-GAN methods provide a promising solution to address the challenges of limited data and computational power in generating dolphin signals.

Efficient data dimensionality reduction method for improving road crack classification algorithms

AbstractAutomatic crack classification plays an essential role in road maintenance. Using many features for the classification is inefficient for implementing embedded systems with low computational resources makes it difficult. Therefore, this work proposes a new data dimensionality reduction (DDR) for crack classification algorithms (DDR4CC). DDR4CC reduces the required information about the cracks to only four features. Using these features, the images can be classified into longitudinal, transverse, and alligator cracks or healthy pavement. DDR4CC is compared with eight DDR methods, and the reduced set of features is analyzed using five different classification algorithms. Besides, five different datasets, generated by a combination of several public datasets, are used. We are proposing a simple DDR method with high interpretability of the data, obtaining very fast computation and high accuracy. Experiments show that DDR4CC enhances the results of the classification algorithms, providing almost perfect classifiers with a minimum computation time.

Building Thunderstorm Resilience in the Hindu Kush Himalaya Region through Probabilistic Forecasts and Satellite Observations

Abstract Some of the most intense thunderstorms on Earth occur in the Hindu Kush Himalaya (HKH) region of southern Asia—where many organizations lack the capacity needed to predict, observe, and/or effectively respond to the threats associated with high-impact convective weather. As a result, a disproportionately large number of casualties and damage often occur with premonsoon severe thunderstorms in this region. To address this problem, we combined ensemble numerical weather prediction (NWP), satellite-based precipitation products, and land-imagery techniques into a High-Impact Weather Assessment Toolkit (HIWAT) customized for HKH. In 2018 and 2019 demonstrations, a regional convection-allowing ensemble NWP system was configured to provide real-time probabilistic guidance of thunderstorm hazards over HKH, applying ensemble techniques developed for U.S.-focused experiments. Case studies of damaging wind, large hail, lightning, a rare Nepalese tornado, and landfalling tropical cyclone events show how HIWAT efficiently packages ensemble output into products that are readily interpreted by forecasters in HKH. Precipitation and total lightning flash verification reveal the highest skill occurred where deep convection was most frequently observed in Bangladesh and northeastern India, and verification scores exceeded global ensemble scores for heavy precipitation rates. These results demonstrate that plausible forecasts of thunderstorm hazards can be attained with relatively low computational resources, thereby facilitating advancements in extreme weather forecasting services in historically underserved regions such as HKH. In early 2022, a custom version of HIWAT was installed at the Bangladesh Meteorological Department using in-house computational resources, providing regional ensemble forecast guidance in real time.

Using Chaos-Game-Representation for Analysing the SARS-CoV-2 Lineages, Newly Emerging Strains and Recombinants.

Viruses have high mutation rates, facilitating rapid evolution and the emergence of new species, subspecies, strains and recombinant forms. Accurate classification of these forms is crucial for understanding viral evolution and developing therapeutic applications. Phylogenetic classification is typically performed by analyzing molecular differences at the genomic and sub-genomic levels. This involves aligning homologous proteins or genes. However, there is growing interest in developing alignment-free methods for whole-genome comparisons that are computationally efficient. Here we elaborate on the Chaos Game Representation (CGR) method, based on concepts of statistical physics and free of sequence alignment assumptions. We adopt the CGR method for classification of the closely related clades/lineages A and B of the SARS-Corona virus 2019 (SARS-CoV-2), which is one of the fastest evolving viruses. Our study shows that the CGR approach can easily yield the SARS-CoV-2 phylogeny from the available whole genomes of lineage A and lineage B sequences. It also shows an accurate classification of eight different strains and the newly evolved XBB variant from its parental strains. Compared to alignment-based methods (Neighbour-Joining and Maximum Likelihood), the CGR method requires low computational resources, is fast and accurate for long sequences, and, being a K-mer based approach, allows simultaneous comparison of a large number of closely-related sequences of different sizes. Further, we developed an R pipeline CGRphylo, available on GitHub, which integrates the CGR module with various other R packages to create phylogenetic trees and visualize them. Our findings demonstrate the efficacy of the CGR method for accurate classification and tracking of rapidly evolving viruses, offering valuable insights into the evolution and emergence of new SARS-CoV-2 strains and recombinants.

Three-axis high-accuracy spacecraft attitude estimation via sequential extended Kalman filtering of single-axis magnetometer measurements

Magnetometer is a highly advantageous sensor for determining a spacecraft’s attitude. This article provides a solution to the problem of spacecraft attitude estimation using magnetometer measurements only. To ensure full observability of spacecraft attitude states, it is necessary to use at least two types of sensors. Consequently, utilizing a single sensor, such as the magnetometer, poses a significant challenge for any attitude estimation algorithm, including the extended Kalman filter (EKF). Moreover, implementing the EKF algorithm, or any other attitude estimation algorithm, is computationally intensive. To address these issues, an algorithm has been developed that estimates spacecraft attitude angles and attitude rates using a sequential extended Kalman filter (SEKF). This algorithm offers numerous benefits over those found in the literature such as high accuracy, low computational resource requirements, the ability to converge even with large initial attitude and angular velocity estimation errors, and the ability to function even if two of the three measurement channels of the magnetometer are not functioning. With these benefits, the developed SEKF algorithm is capable of operating in all spacecraft operational modes, delivering accurate performance and computation time. In spite of measurements with large noise values, the high accuracy achieved by the SEKF algorithm enables the magnetometer to serve as the sole source of attitude information, even if one or two magnetometer measurement channels are not functioning.

Nonlinear autoregressive with external input neural network for predicting the nonlinear dynamics of supercontinuum generation in optical fibers

Full characterization of the propagation dynamics of ultra-short pulses in optical fibers is of fundamental importance in designing optical devices for several applications in the nonlinear optics field. Such applications range from basic descriptions of the light–matter dynamics to Bose–Einstein condensates, plasma physics, hydrodynamics, high-resolution imaging, and remote sensing, among many others. Nevertheless, ultra-short pulse propagation is a highly nonlinear process, so correctly describing all temporal and spectral features of these pulses is a big challenge, consuming extensive computational resources. Looking for simple solutions to this problem, we present in this paper, for the first time, to the best of our knowledge, a nonlinear autoregressive with external input neural network (NARXNET) capable of predicting the nonlinear dynamics of supercontinuum generation in optical fibers. The NARXNET structure allows low prediction error, fast training as short as 1.45 min, satisfactory generalization ability, and low computational resources for the training and testing stages.

Marine debris detection using a multi-feature pyramid network

ABSTRACT Marine Debris is the submerged and floating litter that majorly consists of plastics or other manmade polymers that are either intentionally discarded or unknowingly lost along shores of beaches, rivers or other waterbodies. Significant water pollution is caused due to such untreated waste released into the oceans, specifically in the coastal regions. The recent years have seen an immense rise in marine debris, which has created a necessity to reduce the amount of human effort and led to an increase in the demand for automatic marine debris detection. This paper proposes a multi-feature pyramid network (MFPN) of three subnetworks feature extractor, feature pyramid and pooling block that are concatenated to form a single end-to-end network for the task of automatic marine debris detection from multispectral imagery. Further, the feature extractor network is initialized with the pre-trained weights of Resnet-101 to achieve high performance with lower computational resources and normalization is used to reduce the bias of the model. The proposed model is tested on the Sentinel-2-based marine debris benchmark dataset MARIDA and outperforms the state-of-the-art architectures in the field by reporting an mIOU, Pixel Accuracy and F1-Score of 0.71, 0.84 and 0.80, respectively.

Hybrid Extend Particle Swarm Optimization (EPSO) model for Enhancing the performance of MANET Routing Protocols

The routing protocols in MANETs are designed to provide efficient and reliable communication in a highly dynamic and resource-constrained environment. It is very efficient and requires low computational and memory resources compared to most routing protocols. Therefore, mobility and the number of nodes significantly impact the performance and reliability of routing protocols. This paper proposes a hybrid extended particle swarm optimization (EPSO) model to improve the performance of MANET routing protocols. It determines the optimal mobility and the number of hubs and nodes that satisfy the best possible version of MANET. MANET requires a robust routing algorithm that can adapt to a network that arbitrarily changes its topology at any time. The proposed model in the NS2 simulator proves the model's validity in improving the performance of MANET. The proposed model sets the general parameters of routing protocols and achieves high performance with fewer discarded packets and low delay when sending and receiving over MANET. The MANET sent 167 packets in the proposed model, and the number of discarded packets was less than 1%.

Intelligent identification of radar active jamming type based on multi-domain information fusion

Feature fusion is beneficial to improving the recognition rate of radar active jamming types, which usually costs a lot of computation and network complexity, however. Contrapose the disadvantage of low computational power and resource allocation of missile-borne radar, this paper introduces data enhancement theory into the feature fusion method and proposes an intelligent identification model of radar active jamming type based on multi-domain information fusion. The model first analyzes the time-frequency domain features of the signal and one-dimensional features such as its real part, imaginary part, frequency spectrum, and power spectrum, and then uses the Cutout&Patchup algorithm to fuse the one-dimensional and two-dimensional features into a new multi-domain information fusion matrix as the input of the recognition network. The simulation results show that this method greatly improves the recognition accuracy of active jamming types. Under the WideResNet28_2 classifier, the classification accuracy of active interference type reaches 88.06%, which is 0.79% higher than that before fusion.

Sensitivity analysis for forecasting Brazilian electricity demand using artificial neural networks and hybrid models based on Autoregressive Integrated Moving Average

Several studies focus on improving forecasting techniques to capture multiple patterns in time series. The evolution of computing hardware has made possible to solve complex equations with large amount of data, such as the one used in neural networks. On the other hand, time series methods such as ARIMA (Autoregressive Integrated Moving Average) could also have a good approximation with low computational resources. Nonetheless, to improve the ARIMA approximations, it could be combined with other techniques such as Wavelet Transform or Fourier Transform. Therefore, this work evaluates the appropriate utilization to make predictions for different time horizons (2, 5 and 10 years) and different time frequencies (days, months, and years) using artificial neural network, ARIMA combined with Wavelet Transform, or Fourier Transform. The results show that Artificial Neural Networks provides a better approach for short-term horizons considering either time frequency, ARIMA with Fourier Transform has the best approximation for the monthly time series and either time horizons and ARIMA with Wavelet Transform has the best approximation for medium-term and long-term periods with either time frequency.

ECG data compression using the formation of QRS-complex segment bank and integer DCT-based plateau region processing

Electrocardiogram (ECG) data is essential for evaluating cardiac health conditions. Long-term ECG signal monitoring needs a considerable amount of memory for storage, impacting data bandwidth allocation. The ECG data compression is challenging in preserving significant features, using low computational resources, and achieving a higher compression ratio at lower reconstruction error. This paper proposes an innovative QRS-complex and plateau (Non-QRS) region segregation for QRS-complex subsegment bank formation and integer cosine transform-based ECG data compression. An adaptive QRS-complex subsegment template bank is formed by: (i) QRS and Non-QRS region segmentation; (ii) QRS region identification; (iii) QRS region alignment; (iv) Dc equalization and sub-template formation of the QRS region. A discrete integer cosine transform, quantization, and grouping with a significance map are used on the non-QRS segment of the ECG. Both the QRS and non-QRS regions are coded by lossless arithmetic encoding. Experiments are enumerated on the MIT BIH arrhythmia database, which validates the potential of the proposed approach by achieving an average value equal to 31.51, 0.84, and 28.57 for compression ratio, percentage-root-mean-square-difference, and quality-score, respectively. The average distortion measure for the proposed ECG compressor is less than 2%, which is within the clinically acceptable limit. The proposed method is also compared to existing state-of-the-art ECG compression techniques. It performs better in terms of compression ratio and preservation of ECG features. Additionally, this method typically requires less computational resources and is easy to implement, which is an advantage over other ECG compression methods.

Eulerian vs Lagrangian Method for Low Computational Resources: A Comparison of 2D Dam Break Case

It is well known that computational fluid dynamics has been challenging for low-computational-resourced students and researchers. In this study, we performed a 2D dam break simulation by Eulerian approach-finite volume method of OpenFOAM and Lagrangian approach-finite volume particle method (FVP) in a low-specification personal computer. We compared those approaches’ results qualitatively and quantitatively against experimental data and measured their simulation time for 6 different grid sizes. We compared the visualization results and their pressure and velocity using ParaView and VisIt. At the same time, a quantitative comparison was made by determining the waterfront position each time using the Tracker video analyzer. It was found that OpenFOAM resulted in better visualization results, lower error by 28 - 68 %, and more reasonable pressure and velocity profiles. It can also simulate smaller grid sizes and 10 to 200 times faster than FVP, but it can’t produce the air-liquid interface’s sharpness as good as FVP. Thus, to simulate cases where interfaces are not important, Eulerian-based OpenFOAM is best suited to perform in a low-specification computer. HIGHLIGHTS One of the main challenge in CFD is that it needs a huge computational resources to get optimum results. For students or researchers with limited computational resources, which approaches can they used? Two approaches in CFD: Eulerian and Lagrangian methods-represented by OpenFOAM and Finite Volume Particle Method (FVP)-were performed in a personal laptop to simulate 2D dam break problem and their results are compared with the experimental data Qualitative and quantitative comparison shows that OpenFOAM is well suited for low computational resources users. However, FVP method has an advantage in terms of capturing the interface of the two phases involved in simulation. Thus, for phase-change phenomena where interface is important, it is strongly recommended to use FVP-though one might need a more sophisticated resources than a personal laptop GRAPHICAL ABSTRACT

Timestamp Based OTP and Enhanced RSA Key Exchange Scheme with SIT Encryption to Secure IoT Devices

The Internet of Things (IoT) has become an emerging technology and is expected to connect billions of more devices to the internet in the near future. With time, more and more devices like wearables, intelligent home systems, and industrial automation devices are getting connected to the internet. IoT devices primarily transfer data using wireless communication networks, introducing more vulnerabilities like man-in-the-middle-attacks and eavesdropping. These security concerns are customary for any device communicating over the internet because of its intrinsic open nature. These problems are usually subdued by conventional cryptographic algorithms used in typical systems that are power-hungry and computationally intensive, making them infeasible to be used in IoT devices since they run on low-powered chips, limiting performance, memory, and bandwidth. Hence, there is a requirement to adopt lightweight cryptographic algorithms that can abate the security issues while using low computational resources, which is the constraint in the given scenario. Hence, we propose an end-to-end secured IoT system that ensures the system’s integrity is never compromised using lightweight cryptographic algorithms. We propose a three-module system, where the first module handles user authentication using a time-based one-time password, the second secures communication using lightweight enhanced RSA, and the third performs data encryption using Feistel-based enhanced SIT. This kind of system is designed to deal with security challenges in IoT devices, ensuring adequate data security while reducing the computational footprint using lightweight cryptography.

Priority Scheduling Implementation for Exam Schedule

Scheduling is a common problem that has been raised for a long time. Many algorithms have been created for this problem. Some algorithms offer flexibility in terms of constraints and complex operations. Because of that complexity, many algorithms will need huge computation resources and execution time. A platform like a web application has many restrictions such as execution time and computation resources. A complex algorithm is not suited for the web application platform. Priority scheduling is a scheduling algorithm based on a priority queue. Every schedule slot will produce a queue based on the constraints. Each constraint will have a different weight. Weight in queue represents their priority. This algorithm provides a light algorithm that only needs a few computations and execution times. The exam schedule is one of many problems in educational institutions. A web application is a popular platform that can be accessed from everywhere. Many educational institutions use web platforms as their main system platform. Web platforms have some restrictions such as execution time. Due to web platform restrictions, priority scheduling is a suitable algorithm for this platform. In this study, the author tries to implement a priority scheduling algorithm in scheduling cases with a website platform and shows that this algorithm solution can be an alternative for solving scheduling cases with low computational resources.

Affordable de novo generation of fish mitogenomes using amplification-free enrichment of mitochondrial DNA and deep sequencing of long fragments.

Biomonitoring surveys make use of metabarcoding tools to describe the community composition. These studies match their sequencing results against public genomic databases to identify the species. However, mitochondrial genomic reference data are yet incomplete, only a few genes may be available, or the suitability of existing sequence data is suboptimal for species level resolution. Here, we present a dedicated and cost-effective workflow with no DNA amplification for generating complete fish mitogenomes for the purpose of strengthening fish mitochondrial databases. Two different strategies using long fragment sequencing with Oxford Nanopore technology coupled with mitochondrial DNA enrichment were used. One where the enrichment is achieved by preferential isolation of mitochondria followed by DNA extraction and nuclear DNA depletion ("mitoenrichment"). A second enrichment approach takes advantage of the CRISPR Cas9 targeted scission on previously dephosphorylated DNA ("targeted mitosequencing"). The sequencing results varied between tissue, species, and integrity of the DNA. The mitoenrichment method yielded 0.17%-12.33% of sequences on target and a mean coverage ranging from 74.9 to 805-fold. The targeted mitosequencing experiment from native genomic DNA yielded 1.83%-55% of sequences on target and a 38 to 2123-fold mean coverage. These produced complete mitogenomes of species with homopolymeric regions, tandem repeats, and gene rearrangements. We demonstrate that deep sequencing of long fragments of native fish DNA can be achieved with low computational resources in a cost-effective manner, opening the discovery of mitogenomes of nonmodel or understudied fish taxa to a broad range of laboratories worldwide.

An efficient model to estimate the soil profile and stratigraphic uncertainty quantification

Comprehensive geological information is very important and needed in underground construction and underwater infrastructure. However, it is quite difficult to reveal the configuration of different soil features of underground with sparsely measured data. In this study, a new algorithm named General Regression-Markov Chain (GR-MC) was proposed by coupling the general regression neural network and the Markov chain method together to reproduce the characteristics of underground stratification based on rare borehole information. The general regression neural network method was introduced to evaluate the geological variation along the inclined orientation, while the Markov chain method was applied to examine the change in soil types along the depth. The spatial coordinates of the treated construction site were designed as the input variables, while a vector including the probability information of different soil types was set as the output. The proposed GR-MC method is very intuitive and can be efficiently implemented to predict the geological model using only the coordinates as inputs. Moreover, the proposed method can be smoothly extended into a three-dimensional case with low computational resources. Compared to the existing methods, the proposed GR-MC algorithm is computationally efficient and robust. Three cases from real practice were evaluated and the calculated results were consistent with the known borehole information. Besides, the stratigraphic uncertainty was quantified based on information entropy theory. The results can visually reveal the zones of the estimated soil profile with relatively large uncertainty. The practical application is that possible additional borehole locations were identified to reduce the stratigraphic uncertainty of a construction site.

An intelligent model for supporting edge migration for virtual function chains in next generation internet of things

The developments on next generation IoT sensing devices, with the advances on their low power computational capabilities and high speed networking has led to the introduction of the edge computing paradigm. Within an edge cloud environment, services may generate and consume data locally, without involving cloud computing infrastructures. Aiming to tackle the low computational resources of the IoT nodes, Virtual-Function-Chain has been proposed as an intelligent distribution model for exploiting the maximum of the computational power at the edge, thus enabling the support of demanding services. An intelligent migration model with the capacity to support Virtual-Function-Chains is introduced in this work. According to this model, migration at the edge can support individual features of a Virtual-Function-Chain. First, auto-healing can be implemented with cold migrations, if a Virtual Function fails unexpectedly. Second, a Quality of Service monitoring model can trigger live migrations, aiming to avoid edge devices overload. The evaluation studies of the proposed model revealed that it has the capacity to increase the robustness of an edge-based service on low-powered IoT devices. Finally, comparison with similar frameworks, like Kubernetes, showed that the migration model can effectively react on edge network fluctuations.