Search

This work presents the use of a shallow feedforward artificial neural network (ANN) to develop a prediction model for geometrical deviations in the dry turning of the AA7075 (Al-Zn) alloy. The study focuses on the influence of cutting speed and feed on the arithmetic mean roughness, straightness, and circular runout of cylindrical specimens. The main novelty of this ANN-based model compared to traditional models lies in the simultaneous consideration of geometrical variables at macro and micro scales. The analysis showed that feed was the most influential variable, particularly at higher values, whereas cutting speed had a lesser impact. For all three analysed output variables, the optimal results were achieved by combining low feed and high cutting speed values. The proposed ANN model showed a reasonable adjusted R2 value for all the variables, ranging from 0.87 to 0.97. The ANN performance was compared with other regression models, providing a better fit to the experimental data for all the output variables analysed. Testing of the ANN on additional data not included in the training and validation set confirmed its practical usefulness for predicting geometrical deviations under the studied cutting conditions., CC BY-NC-ND 4.0This work has received funding from the Ministerio de Ciencia e Innovacíon (Gobierno de España), through the research project “Expert system for improving surface integrity in sustainable machining of light alloys (SPAREMETAL)”, with reference PID2021-125988OB-I00. The authors thank the Universidad de Málaga for their contribution to this work, which has received funding from the “II Plan Propio de Investigacíon, Transferencia y Divulgación Científica” of the Universidad de Málaga, through a grant for a research internship at the School of Engineering Science of the University of Skövde (Sweden). Funding for open access charge: Universidad de Málaga / CBUA.

Determining the right selling price for a car can be a challenge for car sales companies. The selling price of a car is highly influenced by car characteristics such as brand, type, year of production, fuel type, and mileage. Therefore, the research aims to develop a more accurate model of car price prediction model by using a stacking ensemble technique that combines Random Forest and ANN. Random Forest is effective in handling outliers and reducing the risk of overfitting, while ANN has the advantage of capturing complex nonlinear patterns. The results show that the stacking ensemble model combining ANN and Random Forest can predict car sales prices by achieving an R2 value of 0.97. The results of this study can help distributors in selling cars make the right decisions regarding the sales price of cars. To improve the generalization of the model, future research is recommended to try a combination of different ensemble methods and the use of larger and more diverse datasets.

The work is devoted to the development of an intelligent system for generating camouflage patterns based on artificial intelligence technologies. A generative-competitive network is used as an intellectual element of this system. To solve the problem of the collapse mode, the architecture of progressively growing GANs (ProGAN) is used. The system allows you to generate completely new camouflage patterns for the selected area by iteratively improving the pattern. Due to the mechanism of restrictions, it is possible to fix the desired aspects of the drawing (color scheme, pattern, number of colors) from an existing drawing and adapt it to the desired area. The system provides the possibility of generating micropatterns on the drawings to improve camouflage at close distances. When evaluating a camouflage pattern, the system takes into account additional parameters, such as angle (from the ground and air), time and weather.

The modern development of industry brings with it an increasing energy demand. In Serbia, a significant portion of electricity is generated by burning coal in thermal power plants, which results in the production of large amounts of industrial waste: fly ash, bottom ash, and slag. These waste materials can contain both toxic and strategically important elements. By determining the quantity of toxic elements present, proper handling of this hazardous waste can be ensured. Additionally, extracting strategic elements from waste materials could offer an ecological alternative to mining. In this study, a comparison of two digestion methods was conducted: HNO3/HCl/HF and HNO3/H2SO4 with V2O5. Additionally, modeling of microwave digestion parameters was performed, including temperature, time, catalyst, and type of mineral acids. The investigations were carried out on seven different fly ash samples. The obtained results were processed using advanced techniques of multivariate chemometric analysis: artificial neural networks (ANN), cluster analysis (CA) and principal component analysis (PCA). The results indicated that lower temperatures and the method HNO3/H2SO4 with V2O5 are more suitable for volatile toxic and some strategic elements, whereas higher temperatures and the method HNO3/HCl/HF are more suitable for other elements, including most of the strategic elements.

Proportional-Integral-Derivative (PID) controllers are extensively used in industrial control applications due to their simplicity and effectiveness in various control tasks. In recent years, there has been a growing emphasis on the integration of Artificial Neural Networks (ANNs) with control theory. This integration aims to harness the versatility of ANNs to facilitate controller design in dynamic environments. Additionally, it opens the possibility of transfer learning between these models, allowing knowledge gained from one control scenario to be applied to others, thereby enhancing adaptability and efficiency in control applications. This thesis explores the potential of ANN structures to replace traditional PID controllers in industrial control systems, focusing on transferability metrics and novel training methodologies. This research presents and analyzes two new transferability metrics, TSM_MI and TSM_COR, for distinct industrial scenarios. TSM_MI, based on mutual information, is effective with existing PID controllers, while TSM_COR's design, focusing on feature and time correlation, suits scenarios without functional target controllers. A novel training framework using scheduled sampling, an input-to-sequence text generation technique, was implemented to address the disparity between training and inference behaviours when developing ANNs functioning as PID controllers, improving metrics such as Integral of Absolute Error (IAE) and Integral of Squared Error (ISE). The findings indicate that the proposed metrics outperform existing methods in specific scenarios, providing a promising measure of ANN transferability. This suggests that ANNs can enhance efficiency and adaptability in various control systems, offering practical solutions and opening lines for further research in ANN-based control strategies.

It is well-known that the training of Deep Neural Networks (DNN) can be formalized in the language of optimal control. In this context, this paper leverages classical turnpike properties of optimal control problems to attempt a quantifiable answer to the question of how many layers should be considered in a DNN. The underlying assumption is that the number of neurons per layer—i.e., the width of the DNN—is kept constant. Pursuing a different route than the classical analysis of approximation properties of sigmoidal functions, we prove explicit bounds on the required depths of DNNs based on asymptotic reachability assumptions and a dissipativity-inducing choice of the regularization terms in the training problem. Numerical results obtained for the two spiral task data set for classification indicate that the proposed constructive estimates can provide non-conservative depth bounds.

Malgrat l'alt rendiment de les Màquines de Vectors de Suport (SVM) en tasques de classificació i regressió (Shawe-Taylor and Cristianini, 2004), les càrregues de dades modernes han introduït desafiaments d'eficiència per a les SVM (Cervantes et al., 2020). Les SVM no són capaces d'encarar problemes no lineals quan la mida del conjunt de dades és considerablement gran. D'altra banda, les arquitectures profundes, com les Xarxes Neuronals Artificials (ANN), han mostrat resultats prometedors al tractar amb patrons complexos en les dades (LeCun et al., 2015). Aprofitant tècniques d'aproximació de funcions kernel com les Random Fourier Features (RFF) (Rahimi and Recht, 2007), aquest treball introdueix una arquitectura de Màquina de Vectors de Suport Multicapa (MLSVM). La MLSVM ha obtingut resultats motivadors, presentant un rendiment competitiu davant de les SVM tradicionals i els mètodes d'última generació. Aquesta tesi obre un camí per a futures investigacions sobre arquitectures de múltiples capes que extenguin el model SVM., Despite the traditional high performance of Support Vector Machines (SVMs) in classification and regression tasks (Shawe-Taylor and Cristianini, 2004), modern data loads have introduced efficiency challenges to SMVs (Cervantes et al., 2020). SVMs are not capable of handling non-linear problems when the dataset size is considerably large. On the other hand, deep architectures such as Artificial Neural Networks (ANNs) have shown encouraging results when dealing with complex patterns in data (LeCun et al., 2015). By leveraging kernel approximation techniques such as Random Fourier Features (RFFs) (Rahimi and Recht, 2007), this work introduces a Multi-Layer Support Vector Machine (MLSVM) architecture. The MLSVM has revealed very promising results, presenting competitive performance against traditional SVMs and state-of-the-art methods. This thesis opens a path for future research on multi-layer architectures extending the SVM model.

The problem of calculating the level of forest cover is considered, including forecasting changes in forest cover in individual forestry. It is stated that the authors previously developed software for calculating forest cover and processing information about forest plantations using the example of the village of Spivakivka in the Izyum district of the Kharkiv region. A comparison of forest cover over a number of years was also made using the Global Forest Watch resource. From this resource, images of Prydonetsk Forestry were taken with conventional designations: areas where new forest plantations are being planted are shown in blue, and areas where cutting is taking place are shown in pink. It is proposed to divide each of the uploaded images of the selected forestry into squares, and then analyze the data for each square. The pink color saturation was calculated and stored in the table. It is noted that forecasting the change in forest stands on the selected site, that is, the change in the percentage of felling, can be done in different ways. First, use regression analysis - apply the regression equation separately to the values of each square, as well as to the entire forestry. Secondly, to form a list of input factors containing indicators on the selected plot in the two previous years and the same indicators on neighboring plots. Thus, the number of factors will be equal to 27: 26 input and 1 output (values on the studied square). Such a forecasting problem can be solved either by the method of multivariate linear regression or by the method of artificial neural networks. The R programming and data analysis language was used to perform calculations using both methods. A script was created that performs calculations by constructing regression lines and an artificial neural network, and also allows determining the best architecture of a neural network and a more effective method of its training for a certain data set. The calculation of the felling dynamics in t

Optimization is a complex process of defining a set of solutions for a wide range of problems, including management decision-making. One of the approaches to increasing the efficiency of solving optimization problems is metaheuristic algorithms. The problem solved in the study is to increase decision-making efficiency in the problems of assessing the state of hierarchical systems while ensuring a given reliability, regardless of its hierarchy. The object of the study is hierarchical systems. The subject of the study is the decision-making process in management problems using an advanced Tasmanian devil algorithm (TDA) and evolving artificial neural networks. A methodical approach using a metaheuristic algorithm is proposed. For TDA training, evolving artificial neural networks are used. The originality of the proposed method lies in setting TDA taking into account the uncertainty of the initial data, improved global and local search procedures. Also, the originality of the study lies in determining TDA feeding locations, which allows prioritizing the search in a given direction. The next original element of the study is the possibility of choosing a TDA hunting strategy, which allows a rational use of available system computing resources. Another original element of the study is determining the initial velocity of each TDA. This makes it possible to optimize the speed of exploration by each TDA in a specific direction. Using the methodical approach provides a 14–17 % increase in data processing efficiency by using additional improved procedures. The proposed methodical approach should be used to solve the problems of evaluating hierarchical systems

The problems of processing heterogeneous data are discontinuous, undifferentiated, and multimodal. The most common approaches to processing heterogeneous data are swarm intelligence algorithms (swarm algorithms). Given the above, classical gradient deterministic algorithms are inappropriate for solving the problems of processing heterogeneous data. The problem solved in the study is to increase the efficiency of processing heterogeneous data circulating in information systems, regardless of the number of data sources. The object of the study is hierarchical systems. A method for increasing the efficiency of processing heterogeneous data using a metaheuristic algorithm is proposed. The study is based on the reptile algorithm (RA) for processing heterogeneous data circulating in the system. For RA training, evolving artificial neural networks are used. The originality of the proposed method lies in setting RA taking into account the uncertainty of the initial data, improved global and local search procedures. Also, the originality of the study lies in determining RA feeding locations, which allows prioritizing the search in a given direction. The next element in the originality of the study is the possibility of choosing an RA hunting strategy, which allows a rational use of available system computing resources. Another original element of the study is determining the initial velocity of each RA. This makes it possible to optimize the speed of exploration of each RA in a certain direction. The method provides a 15–19 % increase in data processing efficiency by using additional improved procedures. The proposed method should be used in processing large amounts of data

The world has recently witnessed a rapid revolution in multimedia signal processing. Images are one of the most widely used media that needs a large amount of data to be represented. Because of the restrictions of limited bandwidth and storage capacity, image compression is a necessity. AMBTC is a straightforward lossy image compression scheme, and studies are still being conducted to improve its performance. This paper incorporated AMBTC with block classification and artificial neural networks to lower the bitrate and preserve the image quality. The proposed scheme was benchmarked with the recent AMBTC techniques for greyscale images. According to the results, the proposed method significantly improved over conventional AMBTC by achieving 16% bitrate reduction while preserving 99.19% of AMBTC’s Peak Signal to Noise Ratio (PSNR).

The output power of drive should be controlled to avoid stress on the advanced components in the input power system of the 1-∅ AC source; this is powered to the three-phase (3-∅) variable frequency drives (VFDs). To deal these issues, an integrated artificial neural network (ANN) and fuzzy logic control (FCL) named as adaptive neuro fuzzy inference system (ANFIS)-based VFD optimization is proposed to mitigate the stresses above the various parts of VFD such as input side, terminal block, direct current (DC) capacitor bus, current harmonics, torque ripple and speed of the induction motor (IM). In addition, the proposed ANFIS with the supervisory learning approach is utilized to regulate the speed with mitigated rise time and settling time of the VFD system. The proposed ANFIS model is simulated in MATLAB/Simulink environment and compared with several conventional VFD optimization The extensive simulated performance shows that the proposed ANFIS-based VFD has achieved better results than conventional VFD optimization techniques.

Open biomass burning (OBB) is the main cause of air pollution in Northern Thailand, where PM2.5 concentrations exceed Thailand's air quality standards annually during the January–April (dry season). The air emissions from databases that detail the pollutants discharged into the atmosphere from specific sources of air pollution are crucial for monitoring air pollution. However, this data has been poorly studied in Thailand. This study estimated ground-level PM2.5 concentration in Northern Thailand using the Multilayer Perceptron Artificial Neural Networks (MLP-ANN) model, integrating the in-depth data as input variables. The 10-fold cross-validation approach was applied to validate the model's performance. The meteorological and aerosol optical depth (AOD), which were the satellite data detected by a sensor from the Moderate Resolution Imaging Spectroradiometer (MODIS) Terra and Aqua satellites, were used as input variables. One important input variable was the OBB emission, which consists of forest fire and crop residue burning air emissions. The best modeling was found with the optimal architecture networks, with 8-16-1 indicating the lowest mean absolute error (MAE) values at 0.0187 and root mean square error (RMSE) at 0.0282. The model result was observed as an underestimate between the model result and the actual data from the Pollution Control Department (PCD), for which the limitation of the training datasets was the reason. Moreover, this model was also applied to estimates in Khon Kaen, Nakhon Ratchasima, and Ubon Ratchathani provinces. The model results were still underestimated compared data from the PCD. The primary reasons were the difference in the geographical characteristics and air pollution, including the error model. The highlight finding in this study indicated that integrating the input variables as the meteorological data, the AOD, and the OBB emissions processing with MLP-ANN enhances model performance and can be extended to further estimation

One of the challenges in the drying process is decreasing the drying time while preserving the product quality. This work aimed to assess the impact of pulsed electric field (PEF) treatment with varying specific energy levels (15.2–26.8 kJ/kg) in conjunction with a microwave vacuum dryer (operating at energy levels of 100, 200 and 300 W) on the kinetics of drying apple slices (cv. Gravenstein). The findings demonstrated a notable reduction in the moisture ratio with the application of pulsed electric field treatment. Based on the findings, implementing PEF reduced the drying time from 4.2 to 31.4% compared to the untreated sample. Moreover, two mathematical models (viz. Page and Weibull) and two machine learning techniques (viz. artificial neural network and support vector regression) were used to predict the moisture ratio of the dried samples. Page’s and Weibull’s models predicted the moisture ratios with R2 = 0.958 and 0.970, respectively. The optimal topology of machine learning to predict the moisture ratio was derived based on the influential parameters within the artificial neural network (i.e., training algorithm, transfer function and hidden layer neurons) and support vector regression (kernel function). The performance of the artificial neural network (R2 = 0.998, RMSE = 0.038 and MAE = 0.024) surpassed that of support vector regression (R2 = 0.994, RMSE = 0.012 and MAE = 0.009). Overall, the machine learning approach outperformed the mathematical models in terms of performance. Hence, machine learning can be used effectively for both predicting the moisture ratio and facilitating online monitoring and control of the drying processes. Lastly, the attributes of the dried apple slices, including color, mechanical properties and sensory analysis, were evaluated. Drying apple slices using PEF treatment and 100 W of microwave energy not only reduces drying time but also maintains the chemical properties such as the total

Machine learning models have been revolutionizing the manufacturing industry for a decade now. Many powerful models have been developed to solve previously thought not solvable problems. In the era of rapid development of the machine learning field, the selection of optimal models for the considered problem within the Industry 4.0 setting is still a hard question. In the frist part of this work , we provide an overview of the use cases of machine learning for Industry 4.0. Afterwards, we provide details of the most commonly utilized machine learning models. Finally, the experimental evaluation is performed to compare the analyzed models for two different use cases, namely visual inspection and predictive maintenance. Two machine learning models are compared for each task in order to highlight the main differences and experimental results.

This systematic literature review (SLR) explores current state-of-the-art artificial intelligence (AI) methods for forecasting hotel demand. Since revenue management (RM) is crucial for business success in the hotel industry, this study aims to identify state-of-the-art effective AI-based solutions for hotel demand forecasting, including machine learning (ML), deep learning (DP), and artificial neural networks (ANNs). The study conducted an SLR using the PRISMA model and identified 20 papers indexed in Scopus and the Web of Science. It addresses the gaps in the literature on AI-based demand forecasting, highlighting the need for clarity in model specification, understanding the impact of AI on pricing accuracy and financial performance, and the challenges of available data quality and computational expertise. The review concludes that AI technology can significantly improve forecasting accuracy and empower data-driven decisions in hotel management. Additionally, this study discusses the limitations of AI-based demand forecasting, such as the need for high-quality data. It also suggests future research directions for further enhancing AI forecasting techniques in the hospitality industry., Esta revisão sistemática da literatura(RSL)visa identificar os mais recentes métodos de previsão de procura de um hotelutilizando inteligência artificial(IA). Considerando a importância que a gestão da receitatempara o sucesso empresarial naindustriahoteleira, este estudo visaidentificar modelos e métodos eficazes baseados em IA para a previsão da procura, incluindo Machine Learning, Deep Learninge Artificial Neural Networks. Neste estudo foi realizada uma RSL utilizando o modelo PRISMA, tendo sido identificados 20 artigos indexados naScopus e na Web ofScience. No estudo são abordadas as lacunas na literatura sobre a previsão da procura baseada em IA, destacando-se a necessidade de maior clareza na especificação dos modelos; o impacto da IA na precisão do preço e no desempenho financeiro; e os desafios relativos à qualidade dos dados disponíveis e da capacidade computacional. A revisão conclui que a utilização de IA integrada nos modelos de análise pode melhorar significativamente a precisão da previsão, reforçando as decisões orientadas por dados na gestão hoteleira. Além disso, este estudo discute também as limitações associadas à previsão da procura baseada em IA, como a necessidade de dados de elevada qualidade. São sugeridas futuras linhas de investigação visando melhorar ainda mais as técnicas de previsão incluindo IA na indústria hoteleira.

In this work, different methods for fly ash modification were applied to obtain an adsorbent for the efficient removal of selected pharmaceuticals from a multiclass aqueous solution. Morphological and surface properties of the modified fly ash samples were analyzed by scanning electron microscopy, X-ray fluorescence, X-ray diffraction, Fourier transform infrared spectroscopy, and point of zero charge, and the influence of the applied modifications was determined by comparison with the results obtained for unmodified fly ash. Experimental parameters of the adsorption of the pharmaceutical onto the modified fly ash were optimized, and special attention was paid to the influence of different parameters on the adsorption capacities. Multivariate methods of analysis, such as artificial neural networks, applied to the obtained results showed that the contact time, the initial concentration of the pharmaceutical solution, and the pH value had the strongest influence on the adsorption process. Fly ash modified with chitosan and magnetic iron oxide showed the best adsorption properties (removal efficiency above 80% for the majority of the selected pharmaceuticals), and artificial neural networks confirmed its susceptibility to the modeling process.

Thesis (Master: Engineering: Civil Engineering)--Central University of Technology, Evapo-transpiration (ET) is one of the crucial elements of the hydrological cycle which expedites constant precipitation through the process of condensation. The accurate prediction of ET is essential for irrigated agriculture as it informs appropriate planning and contributes positively to the daily supervision of the irrigation scheme. However, because of the limited data in arid and semi-arid regions, which have been used widely in the traditional Penman-Monteith approach, alternative, reliable and more powerful techniques are used to predict ET. South Africa is one of many countries that fall under the semi-arid zones where the degree of evapo-transpiration is more than the rainfall rate. The aim of this study was to predict evapo-transpiration in the Keiskammahoek Irrigation Scheme located in Eastern Cape, South Africa, using three, time series, prediction models, namely, Auto Regressive Integrated Moving Average (ARIMA), Artificial Neural Networks (ANNs) and Hybrid (ARIMA-ANNs). ARIMA and ANNs models have been used mostly over the years to predict the linear and non-linear time series, and the Hybrid model, developed by Zhang was also applied because of its ability to capture both the linear and non-linear time series. The 18 years (2001 to 2018) ET time series data was extracted from Google Earth Engine, using java script, at Keiskammahoek Irrigation Scheme. Prior to the prediction of ET time series, the time series data were analysed to understand the behaviour of the time series. A detailed time analysis of Keiskamma River Streamflow and Sandile Dam monthly volume, which are water supply sources close to the study area, were analysed, using time series analysis methods such as: the Break for Additive Seasonal (BFAS) and Trend, Wavelength Analysis, Wavelet Coherence, Correlation Statistics, Theil-Sen plots, Man-Kendall Test, Sequential Mann-Kendall Test and Multi-Linear Regression Analysis. Furthermore, tele-connection analysis between the satellite-derived

The present study investigated the volatile constituents of Cymbopogon citratus (lemongrass) grown in a greenhouse environment in Serbia, marking the first commercial cultivation of the plant for essential oil production in the region. The essential oils and hydrolates obtained through steam distillation were analyzed via gas chromatography–mass spectrometry (GC-MS), and the resulting chemical data were further processed using chemometric methods. This study applied quantitative structure retention relationship (QSRR) analysis, employing molecular descriptors (MDs) and artificial neural networks (ANNs) to predict the retention indices (RIs) of the compounds. A genetic algorithm (GA) was used to select the most relevant MDs for this predictive modeling. A total of 29 compounds were annotated in the essential oils, with geranial and neral being the dominant components, while 37 compounds were detected in the hydrolates. The ANN models effectively predicted the RIs of both essential oils and hydrolates, demonstrating high statistical accuracy and low prediction errors. This research offers valuable insights into the chemical profile of lemongrass cultivated in temperate conditions and advances QSRR modeling for essential oil analysis.

Multipoint simulation (MPS) algorithms exploit the dependence structure of the target variable at a series of k-nearest points around each simulation location. Most of the algorithms require the usage of a training image (TI) to either estimate and draw realisations from the conditional distribution of the target variable at an unsampled location, or else directly draw from the TI at the earliest occurrence of a sufficiently similar pattern as the conditioning one. The innovation from the TI (i.e. the ability to produce patterns not strictly found in it) happens because of a certain tolerance embedded into these algorithms. We present an MPS that can be trained on extensive conditioning data as well as on a TI. The method has three steps. In a first step, the training data is so to say "folded"" in a pre-specified series of ways to produce long lists of patterns containing (k+1)-tuples of nearest and near observations to each sample location: not just tuples of the target variable are created, but also information about the constellation of the corresponding (k+1) locations of these tuples is added to the patterns. In a second step, the folded dataset is fed into the training of an artificial neural network (ANN) to predict the target variable at the reference point conditional on the surrounding k-point pattern. Once the ANN is trained, in the third step we follow a random path through the simulation grid, each time finding the k-nearest data points and using the trained ANN to predict the conditional target value at each grid node. In this contribution, we discuss the details of the proposed algorith with an illustration example and discuss its properties and applicability.

The results of a study in which deep learning models were developed and implemented for diagnosing heart diseases based on echocardiographic images are presented. Modern diagnostic methods are analyzed, their limitations and shortcomings are clarified, which prompted the search for new approaches. It was found that traditional methods often do not provide sufficient accuracy in detecting heart pathologies, which can lead to errors in patient diagnosis and treatment. The possibilities of using artificial intelligence, in particular neural networks, to automate the process of analyzing echocardiographic images are investigated. The architecture of a neural network has been developed, which has demonstrated high efficiency in detecting various types of heart diseases. The new model has been compared with traditional methods, which confirmed its advantages in the accuracy of segmentation and classification of pathologies. Algorithms have been introduced that allow significantly increasing the sensitivity and specificity of diagnostics, which is important for the timely detection of diseases. The results of the study indicate that the use of the developed model can considerably improve the quality of medical services, reducing the risk of errors in diagnosis and increasing the effectiveness of treatment. The importance of integrating modern technologies into medical practice is emphasized, which opens up new opportunities for early detection of heart diseases and improving the general health of the population. The results of the study can become the basis for further scientific developments in this field, as well as for the implementation of new diagnostic standards in clinical practice. The architecture of the neural network is analyzed, the influence of various parameters, such as the number of epochs and batch size on the learning efficiency is investigated. The results presented confirm the advantages of the new model over traditional methods, and potential direction

The thesis investigates applications and extensions of reinforcement learning (RL) algorithms to modeling human cognition, and focuses on development of new tools for fitting cognitive models to behavioral data. The first part of the thesis examines the effect of choice abstraction on recruitment of RL mechanisms. This work challenges the basic RL assumption that action space is always finite and defined, and tests the variability in processes that best describe the data when the appropriate choice features are ambiguous (e.g. abstract). Results indicate that when choices of multiple levels of abstraction are plausible, less abstract choices (e.g. simple motor actions) interfere with more abstract choices (e.g. goal selection). Further cognitive modeling and experimental tests showed that working memory (WM) contribution to more abstract choice process was reduced relative to that of RL, potentially due to the use of WM resources for defining the appropriate choice features in the abstract condition. Second project explored the effect of subgoals, the intermediate learning milestones, on learning in the context of hierarchical reinforcement learning (HRL) framework. In this project we operationalized subgoals in a way that removes the features commonly associated with subgoals (novelty, reward associations, frequency) and sought to test whether subgoals contribute to learning hierarchically organized policies and generalization through a pseudoreinforcing effect independent of these features. The results revealed that participants solved the hierarchical task, with data patterns implying the effect of subgoals on behavior; generalization tests showed that generalization of subgoals, under the constraint of our subgoal definition, was possible but predicated on explicit recognition of subgoal features. The third project focused on development of new cognitive model-fitting tool leveraging artificial neural networks (ANN). The results demonstrating ANN efficacy in fit

The accelerated development of applications related to artificial intelligence has generated the creation of increasingly complex neural network models with enormous amounts of parameters, currently reaching up to trillions of parameters. Therefore, it makes your training almost impossible without the parallelization of training. Parallelism applied with different approaches is the mechanism that has been used to solve the problem of training on a large scale. This paper presents a glimpse of the state of the art related to parallelism in deep learning training from multiple points of view. The topics of pipeline parallelism, hybrid parallelism, mixture-of-experts and auto-parallelism are addressed in this study, which currently play a leading role in scientific research related to this area. Finally, we develop a series of experiments with data parallelism and model parallelism. The objective is that the reader can observe the performance of two types of parallelism and understand more clearly the approach of each one.

The current popularity in the application of artificial intelligence to solve complex problems is growing. The appearance of chats based on artificial intelligence or natural language processing has generated the creation of increasingly large and sophisticated neural network models, which are the basis of current developments in artificial intelligence. These neural networks can be composed of billions of parameters and their training is not feasible without the application of approaches based on parallelism. This paper focuses on studying pipeline parallelism, which is one of the most important types of parallelism used to train neural network models in deep learning. In this study we offer a look at the most important concepts related to the topic and we present a detailed analysis of 3 pipeline parallelism libraries: Torchgpipe, FairScale, and DeepSpeed. We analyze important aspects of these libraries such as their implementation and features. In addition, we evaluated them experimentally, carrying out parallel trainings and taking into account aspects such as the number of stages in the training pipeline and the type of balance.

Myocardial infarction is a serious disease that can be fatal. Early detection and prevention are crucial to improve the prognosis of patients. Dermatoglyphic studies, which examine skin patterns on the palms and fingers, can help in the early detection of people who are predisposed to myocardial infarction. The aim of the study was to determine the dermatoglyphic phenotype of the feet in people with myocardial infarction. The object of the study was foot dermatoglyphs obtained from 60 male subjects (30 patients with myocardial infarction and 30 healthy men without cardiovascular disease as a control group). The dermatoglyphic method and the method of statistical analysis were used. It was found that in patients with MI, in contrast to the control group, there were loops on the I, II and III fingers and arcs on the V fingers; a decrease in the frequency of curls on the third toe of the left foot and an increase in the frequency of this pattern on the second toes of both feet; an increase in the value of the comb count in the area of the tibial edge of the foot, with a decrease in the values near the fibular edge of the foot. Thus, dermatoglyphic examination is a simple, inexpensive and non-invasive method that can help in the early detection of people who are predisposed to myocardial infarction. This may help in the development of new methods of screening and prevention of myocardial infarction, which may lead to an improved prognosis for patients.

In this work, we present the modelling and numerical simulation of a molten glass fluid flow in a furnace melting basin. We first derive a model for a molten glass fluid flow and present numerical simulations based on the finite element method (FEM). We further discuss and validate the results obtained from the simulations by comparing them with experimental results. Finally, we also present a non-intrusive proper orthogonal decomposition (POD) based on artificial neural networks (ANN) to efficiently handle scenarios which require multiple simulations of the fluid flow upon changing parameters of relevant industrial interest. This approach lets us obtain solutions of a complex 3D model, with good accuracy with respect to the FEM solution, yet with negligible associated computational times.

The present study investigated the volatile constituents of Cymbopogon citratus (lemongrass) grown in a greenhouse environment in Serbia, marking the first commercial cultivation of the plant for essential oil production in the region. The essential oils and hydrolates obtained through steam distillation were analyzed via gas chromatography–mass spectrometry (GC-MS), and the resulting chemical data were further processed using chemometric methods. This study applied quantitative structure retention relationship (QSRR) analysis, employing molecular descriptors (MDs) and artificial neural networks (ANNs) to predict the retention indices (RIs) of the compounds. A genetic algorithm (GA) was used to select the most relevant MDs for this predictive modeling. A total of 29 compounds were annotated in the essential oils, with geranial and neral being the dominant components, while 37 compounds were detected in the hydrolates. The ANN models effectively predicted the RIs of both essential oils and hydrolates, demonstrating high statistical accuracy and low prediction errors. This research offers valuable insights into the chemical profile of lemongrass cultivated in temperate conditions and advances QSRR modeling for essential oil analysis. © 2024 by the authors.

Using a Taguchi experimental design, this research focuses on utilizing indigenous bacteria from the Danube River to enhance the self-healing capabilities and structural integrity of cementitious materials. Bacillus licheniformis and Bacillus muralis were used as individual bacterium or in co-culture, with a concentration of 8 logs CFU, while the humidity variation involved testing wet and wet–dry conditions. Additionally, artificial neural network (ANN) modeling of the compressive strength of cement samples results in improvements in compressive strength, particularly under wet–dry conditions. By inducing targeted bacterial activity, the formation of calcium carbonate precipitates was initiated, which effectively sealed formed cracks, thus restoring and even enhancing the material’s strength. In addition to short-term improvements, this study also evaluates long-term improvements, with compressive strength measured over periods extending to 180 days. The results demonstrate sustained self-healing capabilities and strength improvements under varied environmental conditions, emphasizing the potential for long-term application in real-world infrastructure. This study also explores the role of environmental conditions, such as wet and wet–dry cycles, in optimizing the self-healing process, revealing that cyclic exposure conditions further improve the efficiency of strength recovery. The findings suggest that autochthonous bacterial co-cultures can be a viable solution for enhancing the durability and lifespan of concrete structures. This research provides a foundation for further exploration into bio-based self-healing mechanisms and their practical applications in the concrete industry.

Hydro-power is the most widely used form of renewable energy, supplying over half of the demand for electricity in Sweden. Despite being climate friendly, hydro-power plants can disrupt river connectivity which has a negative impact on biodiversity. Atlantic salmon (Salmo salar) is a species that is highly dependent on having unimpeded river systems to migrate from their birthplace in the river to their feeding grounds in the ocean. Different methods exist to monitor the health of salmon populations in rivers where hydro-power is employed. Traditionally, monitoring has been done by catching and tagging the fish which can cause harm to the fish and is a labor-intensive process. An alternative method is to use image analysis to identify individual salmon by their unique melanophore spot patterns. In this thesis, two methods for identifying individual salmon based on images from two camera traps near Stornorrfors hydro-power plant in Umeälven are presented. The first method involves using template-matching of spot patterns to measure the similarity of salmon in different images. The second method implements similarity learning by training a Siamese neural network. The results show that identifying matching images of the same individual salmon between datasets of images taken from different cameras is possible using both methods. For the spot pattern template matching, the accuracy of the spot pattern annotations is a key factor in achieving a high performance. The capability for this method to find matches does not appear to be significantly reduced when the size of the target set is increased. This differs from the capability of the siamese network, which is able to find a higher proportion of all matches between the source set and a small target set, while finding much fewer correct matches when the size of the target set is increased., Vattenkraft är den vanligast förekommande formen av förnybar energi och står för över hälften av Sveriges elproduktion. Trots sin klimatvänlighet kan vattenkraftverk agera som ett hinder i älven vilket får en negativ påverkan på den biologiska mångfalden. Atlantlax (Salmo salar) är en art som är beroende av att ha obehindrade vattensystem för att förflytta sig från deras födelseplats i älven till deras utfodringsplatser i havet. Det finns olika metoder för att övervaka den ekologiska statusen hos laxpopulationer i älvar där vattenkraft används. Traditionellt har detta gjorts genom att fånga och märka fisken, vilket dels kan skada fisken och dels är en arbetsintensiv process. En alternativ metod är att använda bildanalys för att identifiera individuella laxar med hjälp av deras unika melanoforfläckmönster. I denna avhandling presenteras två metoder för att identifiera individuella laxar med hjälp av bilder från kamerafällor nära vattenkraftverket Stornorrfors i Umeälven. Den första metoden använder mallmatchning av fläckmönster för att mäta likheten mellan laxar i olika bilder. Den andra metoden implementerar likhetsinlärning genom att träna ett siamesiskt neuralt nätverk. Resultaten visar att det är möjligt att identifiera matchande bilder av samma individuella lax mellan en källuppsättning och en måluppsättning bestående av bilder tagna från olika kameror med båda dessa metoder. För matchning av fläckmönster är annoteringen av fläckmönstren en viktig faktor för att uppnå ett bra resultat. Förmågan för denna metod att hitta matchningar minskar inte markant när storleken på måluppsättningen ökas. Detta skiljer sig från det siamesiska nätverket, som kan hitta en högre andel av alla matcher mellan källuppsättningen och en liten måluppsättning, samtidigt som mycket färre korrekta matchningar hittas när storleken på måluppsättningen ökas.

This study aimed to address a workflow challenge in the business sector by applied deep learning, Artificial Intelligence (AI) and automation. It focused on the detection and extraction of tabular data found in Annual General Meeting (AGM) minutes; official records summarizing relevant information gathered during annual firm meetings. The findings offer valuable insights for similar applications in the interface between finance and technology. Two distinct state-of-the-art object detection methods were employed and compared. The first incorporating a Convolutional Neural Network (CNN) variant, specialized for high accuracy object detection. The latter method revolving around transformers, theoretically promising enhanced performance and accuracy. For data extraction, OCR techniques and a different set of metrics were used. The transformer-based approach exhibited slightly higher, overall performance and accuracy with an F1-score of 0.914, compared to the 0.888 returned by the CNN approach. The results suggest that employing the final model as an autonomous tool would require human supervision to ensure accurate results. Conclusively, the model should only be viewed as an assistive resource and not a complete and autonomous solution to the problem., Denna studie syftade till att lösa en arbetsflödes utmaning inom affärssektorn genom tillämpad djupinlärning, artificiell intelligens (AI) och automation. Studien fokuserade på identifiering och extrahering av tabelldata från årsstämmoprotokoll; officiella dokument som sammanfattar relevant information från årliga företagsmöten. Resultaten ger värdefulla insikter för liknande appliceringar inom finans och teknologisektorn. Två distinkta djupinlärningsmetoder för objektidentifiering användes och jämfördes. Den första metoden byggde på en variant av ett konvolutionellt neural nätverk (CNN), specialiserat på objektidentifiering med hög noggrannhet. Den andra metoden baserades på transformers, vilka i teorin utlovar förbättrad prestanda och noggrannhet. För dataextraktion användes optiska teckenigenkänningstekniker (OCR) och andra mätvärden. Den transformer-baserade metoden uppvisade högre total prestanda och noggrannhet med en F1-Poäng på 0.914 jämfört med 0.888 för CNN metoden. Resultaten tyder även på att den slutliga modellen skulle kräva mänsklig översyn för att kunna användas som ett autonomt verktyg och för att säkerställa noggranna resultat. Sammanfattningsvis bör modellen betraktas som ett stödverktyg och inte en komplett och autonom lösning till det givna problemet.

Turbulent exchanges between sea ice and the atmosphere are known to influence the melting rate of sea ice, the development of atmospheric circulation anomalies and, potentially, teleconnections between polar and non-polar regions. Large model errors remain in the parametrization of turbulent heat fluxes over sea ice in climate models, resulting in significant uncertainties in projections of future climate. Fluxes are typically calculated using bulk formulae, based on Monin-Obukhov similarity theory, which have shown particular limitations in polar regions. Parametrizations developed specifically for polar conditions (e.g. representing form drag from ridges or melt ponds on sea ice) rely on sparse observations and thus may not be universally applicable. In this study, new data-driven parametrizations have been developed for surface turbulent fluxes of momentum, sensible heat and latent heat in the Arctic. Machine learning has already been used outside the polar regions to provide accurate and computationally inexpensive estimates of surface turbulent fluxes. To investigate the feasibility of this approach in the Arctic, we have fitted neural-network models to a reference dataset (SHEBA). Predictive performance has been tested using data from other observational campaigns. For momentum and sensible heat, performance of the neural networks is found to be comparable to, and in some cases substantially better than, that of a state-of-the-art bulk formulation. These results offer an efficient alternative to the traditional bulk approach in cases where the latter fails, and can serve to inform further physically based developments.

Experimental nuclear structure data coming from superheavy nuclei synthesis experiments often consists of correlated alpha decay chains. In the absence of neutron detectors - which would fully characterize the exit channel after the fusion-evaporation reaction - the sequence of decay energies and half-lives are the 'fingerprint' of the exit channel itself. Experimental data in this region is sparse, and its interpretation can be liable to error or confirmation bias. A so-called "Schmidt test" is a method for determining the congruence of correlation times for a set of measurements of one decay step. Its outcome is not always entirely conclusive, however. This study evaluates the congruence derived from the Schmidt test using Monte Carlo simulated data with various level of contamination from incongruent data. Furthermore, the study also includes the evaluation of congruence of data stemming from single decays and multi-step decay chains. A multi-layer perceptron was trained on extracted features from simulated decay chain sets with one step. The Schmidt test performs well with larger decay sets and when the half-life of the contaminating species is longer than the original species by a factor 5 or 10. However, the test performs poorly in low counting statistics, where few recorded decay times are available. The newly proposed machine learning model outperforms the Schmidt test in certain high statistics scenarios, but also fails when few decay times are available. Its performance is also poor when the half-life of the contaminant is shorter than the original half-life. The learning behaviour of the model is analysed, showing significant contributions from higher statistical moments in training. Future work involves including chain correlations across multiple steps, alpha decay energies, as well as the potential use of alternative machine learning models., Superheavy nuclei are by definition those with more than 103 protons. They are all man-made and are created through the unlikely fusion of two lighter nuclei. After fusion, these new superheavy nuclei are very radioactive and decay following a characteristic sequence, or 'chain' of radioactive decays. This chain serves as a fingerprint for the newly created nucleus. Experimental chain decay data consists of energies and associated decay times which can be used to confirm the production of a nucleus. However, experimental data is sparse due to the extremely low probability of fusion. This, together with experimental limitations, means that uniquely assigning data to a specific nuclear origin is difficult. This thesis explores a so-called "Schmidt test" that is used to determine whether a set of decay times likely originates from one nucleus or not. A machine learning algorithm is also developed with the same goal as the Schmidt test in order to avoid any bias in the data interpretation. Simulations of decay chain data were performed to evaluate the Schmidt test and the algorithm. In certain cases, the machine learning model outperformed the Schmidt test. However, the algorithm fails to correctly predict congruence when decay times of the contaminant are not sufficiently different from the other decay times. Both the Schmidt test and the machine learning algorithm are rather inconclusive when it comes to situations where very few decay times are available in the data. This work gives some insight into the statistical properties of sets of decay times, but more work is needed to add robustness to the determination of the properties that are most meaningful in identifying the originating nuclei of the decay data.

The continuous investigation into the mechanical properties and rehabilitation management of existing pavements is vital for the sustainability and efficiency of transportation systems. The Mississippi Department of Transportation (MDOT) database, containing over 40 million records, necessitates modernized decision-making models that align with current design methods and materials. In collaboration with the University of Mississippi, MDOT's pavement management program aimed to update outdated Markov transition matrices with advanced Artificial Neural Networks (ANNs). This four-year project focuses on dynamic sequential ANNs to develop performance prediction models for flexible, rigid, and composite pavements, utilizing extensive distress data. Initial results for flexible pavements showed promising outcomes, with robust statistical accuracy measures for rigid pavements (JCP and CRCP) and composite pavements, although further data and calibration are recommended. The developed models, accessible through a user-friendly interface, are expected to improve the prioritization of maintenance and rehabilitation resources, achieving significant time and cost savings while enhancing the overall efficiency of MDOT's pavement management practices.