Search

Diamond burnishing is an effective solution to finish a surface. The purpose of the current work is to optimize parameter inputs, including the spindle speed (S), depth of penetration (D), feed rate (f), and diameter of tool-tip (DT) for improving the Vickers hardness (VH) and decreasing the average roughness (Ra) of a new diamond burnishing process. A set of burnishing experiments is executed under a new cooling lubrication system comprising the minimum quantity lubrication and double vortex tubes. The Bayesian regularized feed-forward neural network (BRFFNN) models of the performances are proposed in terms of the inputs. The criteria importance through the inter-criteria correlation (CRITIC) method and non-dominated sorting genetic algorithm based on the grid partitioning (NSGA-G) are applied to compute the weights of responses and find optimality. The optimal outcomes of the S, D, f, and DT were 370 rpm, 0.10 mm, 0.04 mm/rev, and 8 mm, respectively. The improvements in the Ra and VH were 40.7 % and 7.6 %, respectively, as compared to the original parameters. An effective approach combining the BRFFNN, CRITIC, and NSGA-G can be widely utilized to deal with complicated optimization problems. The optimizing results can be employed to enhance the surface properties of the burnished surface.

Modelling Dynamic Systems with Artificial Neural Networks and Related Methods can be used as a textbook for the field it covers or as an introductory textbook for more advanced literature in the field. The textbook comprises introduction to artificial neural networks, identification of linear and nonlinear dynamic systems, control with artificial neural networks, local-model networks and blended multiple-model systems, design of gain-scheduling control and identification of nonlinear systems with Gaussian processes. It is intended for undergraduate students, especially at the postgraduate level, who have sufficient knowledge in dynamic systems, as well as for professionals who wish to familiarise themselves with the concepts and views described. The textbook is not intended to be a detailed theoretically based description of the subject, but rather an overview of the field of identification of dynamic systems with neural networks and related methods from the perspective of systems theory and, in particular, its application. The work is intended to inform the reader about the views on this subject, which are related but treated very differently.

Izraz incel označuje osebo, v veliki večini primerov moškega spola, ki ocenjuje, da je v neprostovoljnem celibatu, in pogosto izraža sovražnost do spolno aktivnih posameznikov. V preteklih letih so se prepričanja incelov s forumov, namenjenih izključno njim (npr. Incels.is), razširila na popularna družabna omrežja (npr. Reddit), zato sem želela med objavami s foruma, ki primarno ni namenjen incelom, identificirati objave, podobne tistim, ki so značilne za incele ter jih razločiti od objav, ki vsebujejo sovražni, toksični ali žaljivi govor in nevtralni govor. To sem storila s pomočjo procesiranja naravnega jezika in globoke nevronske mreže. Napovedni model oz. nevronsko mrežo, ki je vsebovala jezikovni model ELECTRA, sem učila na zapisih s foruma Incels.is in zapisih iz prosto dostopnih podatkovnih setov, ki so vsebovali sovražni, toksični ali žaljivi in nevtralni govor. S pomočjo modela sem napovedovala kategorije oz. vrsto govora v zapisih s foruma Reddit. Z namenom dodatnega preverjanja veljavnosti modela sem na naključnih podvzorcih zapisov trenirala osem modelov z enakimi nastavitvami kot pri prvotnemu ter med seboj primerjala njihove napovedi vrste govora zapisov s foruma Reddit. V obeh korakih sem izvedla analizo smiselnosti napovedi in vsebinsko analizo klasificiranih objav. Model, treniran na celotnem podatkovnem setu, je pravilno klasificiral 64 % objav, med njimi je največji delež pravilnih klasifikacij spadal v kategorijo govora incelov. Parcialni modeli so bili nekoliko manj točni, njihove napovedi pa so bile srednje skladne. Vsebinska analiza je pokazala, da je večina klasifikacij govora incelov in sovražnega, toksičnega ali žaljivega govora nesmiselnih, večina klasifikacij nevtralnega govora pa smiselnih. Smiselno klasificirane objave v kategoriji govora incelov so vsebovale negativna in stereotipna stališča do žensk, tematike, značilne za govor incelov ter izraze težav v duševnem zdravju pri avtorjih objav. Vse omenjene tematike so izjemno pogoste tudi na forumih, namenjenih incelom, kar kaže na potencialno ranljivost uporabnikov foruma Reddit za razvoj ideologije incelov. Na podlagi ugotovitev svoje raziskave sem podala tudi konkretne predloge preventivnih dejavnosti, s katerimi bi lahko pri moških zmanjšali možnost za razvoj ideologije incelov. The term incel describes people, mostly men, who are involuntarily celibate, and are often hateful towards sexually active individuals. In recent years their beliefs have migrated from forums, exclusive to incels (e.g., Incels.is), towards popular social media (e.g., Reddit). Therefore, I used natural language processing to analyze posts from a forum, not exclusive for incels, and identify posts that are similar to those of the incels, and distinguish them from hate, toxic or offensive speech and neutral speech. I developed a predictive model by using an artificial neural network that included ELECTRA language model and trained it on posts from the forum Incels.is and publicly available posts which contained hate, toxic or offensive speech and neutral speech. I used the model to predict speech type categories of posts from Reddit. To additionally assess the validity of the model I established eight partial models with identical settings as the first model and trained them on random subsamples of posts and compared their predictions on Reddit posts. In both steps I evaluated the reasonableness of the predictions and conducted content analysis of the classified posts. The model that was trained on the entire dataset correctly classified 64% of the posts. Majority of correct classifications belonged to the incel speech category. Partial models were less accurate and the agreement between their classifications was moderate. Content analysis of the posts has shown that the classifications of incel, hate, toxic or offensive speech were mostly unreasonable, while the classifications of neutral speech were mostly reasonable. Reasonably classified posts of incel speech contained negative attitudes towards women, themes that are typical for incel speech, and expressions of authors’ mental health issues. These elements are also commonly present on incel forums, which implies that those Reddit users whose posts contain these elements are vulnerable for the development of incel ideology. The findings of my study were used to provide ideas for prevention strategies that could be implemented to prevent the development of incel ideology in men.

A crow search algorithm (CSA) was applied to perform the optimization of a running blade prosthetics (RBP) made of composite materials like carbon fibre layers and cores of acrylonitrile butadiene styrene (ABS). Optimization aims to increase the RBP displacement limited by the Tsai-Wu failure criterion. Both displacement and the Tsai-Wu criterion are predicted using artificial neural networks (ANN) trained with a database constructed from finite element method (FEM) simulations. Three different cases are optimized varying the carbon fibre layers orientations: –45°/45°, 0°/90°, and a case with the two-fibre layer orientations intercalated. Five geometric parameters and a number of carbon fibre layers are selected as design parameters. A sensitivity analysis is performed using the Garzon equation. The best balance between displacement and failure criterion was found with fibre layers oriented at 0°/90°. The optimal candidate with –45°/45° orientation presents higher displacement; however, the Tsai-Wu criterion was less than 0.5 and not suitable for RBP design. The case with intercalated fibres presented a minimal displacement being the stiffer RBP design. The damage concentrates mostly in the zone that contacts the ground. The sensitivity study found that the number of layers and width were the most important design parameters.

The monitoring of seasonal radial growth of woody plants addresses the ultimate question of when, how and why trees grow. Assessing the growth dynamics is important to quantify the effect of environmental drivers and understand how woody species will deal with the ongoing climatic changes. One of the crucial steps in the analyses of seasonal radial growth is to model the dynamics of xylem and phloem formation based on increment measurements on samples taken at relatively short intervals during the growing season. The most common approach is the use of the Gompertz equation, while other approaches, such as general additive models (GAMs) and generalised linear models (GLMs), have also been tested in recent years. For the first time, we explored artificial neural networks with Bayesian regularisation algorithm (BRNNs) and show that this method is easy to use, resistant to overfitting, tends to yield s-shaped curves and is therefore suitable for deriving temporal dynamics of secondary tree growth. We propose two data processing algorithms that allow more flexible fits. The main result of our work is the XPSgrowth() function implemented in the radial Tree Growth (rTG) R package, that can be used to evaluate and compare three modelling approaches: BRNN, GAM and the Gompertz function. The newly developed function, tested on intra-seasonal xylem and phloem formation data, has potential applications in many ecological and environmental disciplines where growth is expressed as a function of time. Different approaches were evaluated in terms of prediction error, while fitted curves were visually compared to derive their main characteristics. Our results suggest that there is no single best fitting method, therefore we recommend testing different fitting methods and selection of the optimal one. Nasl. z nasl. zaslona. Opis vira z dne 21. 7. 2022. Bibliografija: str. 7-8.

There is a growing interest in implantable drug delivery systems (DDS) in pharmaceutical science. The aim of the present study is to investigate whether it is possible to customize drug release from implantable DDSs through drug–carrier interactions. Therefore, a series of chemically similar active ingredients (APIs) was mixed with different matrix-forming materials and was then compressed directly. Compression and dissolution interactions were examined by FT-IR spectroscopy. Regarding the effect of the interactions on drug release kinetics, a custom-made dissolution device designed for implantable systems was used. The data obtained were used to construct models based on artificial neural networks (ANNs) to predict drug dissolution. FT-IR studies confirmed the presence of H-bond-based solid-state interactions that intensified during dissolution. These results confirmed our hypothesis that interactions could significantly affect both the release rate and the amount of the released drug. The efficiencies of the kinetic parameter-based and point-to-point ANN models were also compared, where the results showed that the point-to-point models better handled predictive inaccuracies and provided better overall predictive efficiency.

Zmožnost delovanja (in zmagovanja) v igrah se pri umetni inteligenci pogosto uporablja kot pokazatelj oz. merilo splošnejše sposobnosti. S stopnjevanjem izzivov pa so zaradi tehničnih ovir odmevni podvigi primorani sklepati kompromise - vmesniki simulacijskih okolij so lahko za umetne agente neskladno prirejeni, kar vzbuja negotovosti v primerjavah z ljudmi. Pregled izbranih del na področju globokega spodbujevalnega učenja v realnočasnih strateških igrah poudarja potrebo po novem merilnem okolju, ki z omogočanjem enakovrednejših vmesnikov bolje izpostavlja vlogo strateških elementov in je hkrati primerno za poskuse na porazdeljenih sistemih. Slednje je izvedeno kot skupinska tekmovalna igra, v opisu katere se obravnavajo določeni tehnični in teoretični problemi na primerih posnemovalnega in spodbujevalnega učenja. Capability of acting (and winning) in games is often used in artifcial intelligence as an indicator or measure of more general ability. However, as challenges escalate, notable efforts are forced to compromise due to technical limitations - interfaces of simulated environments can be inconsistently adapted for artifcial agents, which induces uncertainty in comparisons with humans. Review of select works in the feld of deep reinforcement learning in real-time strategy games highlights necessity for a new benchmark environment, which better emphasises the role of strategic elements by enabling more equivalent interfaces and is also suitable for experiments on distributed systems. The latter is realised as a team-based competitive game, in description of which specifc technical and theoretical problems are examined on the cases of imitation and reinforcement learning.

Na področju strojnega učenja se pogosto pojavljajo problemi z množicami visokih razsežnosti, ki pa so zaradi “prekletstva razsežnosti” zahtevni za reševanje. Pri reševanju takih problemov pogosto uporabljamo metode za manjšanje razsežnosti množic. Popularna metoda za manjšanje razsežnosti so samokodirniki, ki so ponavadi zgrajeni iz nevronskih mrež. Slabost nevronskih mrež je, da zahtevajo veliko procesorskega časa in da ima uporabnik zaradi njihove kompleksnosti zelo slab vpogled v njihovo delovanje. Zato želimo v magistrskem delu razviti samokodirnik na osnovi naključnega gozda, ki teh slabosti ne bi imel. Za konstrukcijo samokodirnika iz naključnega gozda izberemo nabor listov, ki skupaj čim bolje opišejo podatkovno množico, in jih združimo v kodirni vektor. Smokodirnik nato primer zakodira na osnovi njegove pripadnosti listom konstruiranega kodirnega vektorja. Za postopek dekodiranja imamo na razpolago dve informaciji: poti v odločitvenih drevesih, ki vodijo do listov v kodirnem vektorju in shranjene napovedi naključnega gozda. Da poiščemo čim boljšo rekonstrukcijo zakodiranih primerov, uporabimo oba podatka. Naš samokodirnik testiramo, da določimo čim boljše nastavitve parametrov, in njegovo natančnost primerjamo s samokodirniki iz nevronskih mrež. Ugotovimo, da je zaenkrat manj natančen od standardnega pristopa, in premislimo možnosti, kako ga lahko v prihodnosti izboljšamo. In the field of machine learning, problems with high-dimensional data sets are common, and difficult to solve due to the “curse of dimensionality”. To solve these problems, we usually apply methods for dimensionality reduction. A popular method for this are autoencoders, which are usually built with neural networks. However, the downside of neural networks is high computation costs of training and their complexity which obscures the user insight into how they work. To address these issues, we aim at developing an autoencoder that is based on random forests and does not have such problems. To construct an autoencoder from a random forest, we select a set of forest leaves, which describe the data set well, and save them into an encoding vector. We use the encoding vector to encode data samples. There are two types of information we can use to decode the data: the decision tree paths leading to leaves in the encoding vector and the saved predictions form the random forest. We combine the two to get the best possible reconstruction of encoded data. We test the constructed autoencoder to tune the parameter settings and evaluate its performance in comparison to neural network autoencoders. We establish that at this point our autoencoder is significantly less accurate compared to common autoencoders and consider the possibilities for upgrading it in the future.

Avtonomni mobilni roboti v sodobnem industrijskem okolju so obkroženi s številnimi premikajočimi objekti, ki jih robot lahko spremlja s pomočjo svojih zaznaval. V tej nalogi smo za namen napovedovanja trajektorij preučili metode za analizo časovnih vrst s poudarkom na uporabi umetnih nevronskih mrež. Ugotovili smo, da se enkoder/dekoder LSTM mreža lahko uspešno nauči periodičnih vzorcev gibanja robota. Z nadgradnjo te arhitekture smo uspeli napovedovati tudi kratkoročne trajektorije, kar smo v praksi realizirali v obliki ROS vozlišča za napovedovanje trajektorij. Autonomous mobile robots in the modern industrial environment are surrounded by numerous moving objects, which the robot is able to track using its sensors. Often the future position of such objects is needed, therefore we examined the usage of time series methods for trajectory prediction with an emphasis on neural network models. We showed that encoder-decoder LSTM model can successfully learn periodic patterns in the movement of a robot. Enhanced version of this architecture was used to predict short-term trajectories, which we implemented in practice as a ROS node for trajectory prediction.

V doktorski disertaciji nas je zanimala ocena nosilnosti armiranobetonskih (AB) okvirnih konstrukcij po požaru, kar smo izvedli v dveh delih. V prvem delu smo predstavili različne neporušne in porušne metode preizkušanja betona, ki so uporabljene v eksperimentalnem delu. V okviru eksperimentalnega dela je bilo izdelanih pet različnih betonskih mešanic z apnenčevim agregatom, ki se med seboj razlikujejo v vodo-cementnem razmerju, vrsti cementa ter količini vode in dodatkov. Betonski preizkušanci so bili po končani negi in sušenju na zraku v električni peči izpostavljeni temperaturam 200 °C, 400 °C, 600 °C oziroma 800 °C in nato ohlajeni na sobno temperaturo. Sledilo je eksperimentalno preizkušanje, pri čemer so bile uporabljene neporušne in porušne metode. Referenčne vrednosti eksperimentalnih meritev so bile določene na skupini preizkušancev, ki niso bili predhodno segrevani. Rezultati neporušnih preizkusov zajemajo določitev hitrosti preleta vzdolžnih ultrazvočnih valov, površinske trdnosti betona, dinamičnega modula elastičnosti in strižnega modula betona, medtem ko porušni preizkusi zajemajo določitev tlačne in upogibne natezne trdnosti ter modula elastičnosti betona. Nato je bilo s statističnimi metodami ugotovljeno, da temperatura statistično značilno vpliva na omenjene eksperimentalne rezultate, s čimer zaznamo spremembe med posameznimi predhodno segretimi preizkušanci. Sledila je ocena mehanskih lastnosti betona po izpostavljenosti povišanim temperaturam, imenovanih tudi preostale mehanske lastnosti, z regresijskima modeloma z eksplicitnimi zvezami in umetnimi nevronskimi mrežami. Pri tem smo ugotovili, da preostalo upogibno natezno trdnost in modul elastičnosti betona zelo natančno lahko ocenimo na podlagi regresijskih modelov z eksplicitnimi zvezami, medtem ko je za natančnejšo oceno preostale tlačne trdnosti potrebna uporaba umetnih nevronskih mrež. V drugem delu je na kratko predstavljen numerični model za določitev požarne odpornosti linijskih AB konstrukcij po požaru Nfira, ki deluje v programskem okolju Matlab. Novost numeričnega modela so eksperimentalno določeni materialni parametri konstitucijske zveze betona z apnenčevim agregatom po izpostavljenosti povišanim temperaturam. Sledila je izdelava parametričnih študij, pri čemer je bil raziskan vpliv različnih razvojev temperature po požarnem prostoru kot tudi vpliv sestave betonske mešanice na odziv linijskih AB konstrukcij po požaru. In our dissertation we dealt with the estimation of the load bearing capacity of reinforced concrete (RC) frame structures after fire, which we carried out in two parts. In the first part we presented various non-destructive and destructive methods for concrete testing used in the experimental part. Within the experimental investigation we prepared five different concrete mixtures with limestone aggregate, which differ in water to cement ratio, type of cement and the amount of water and additives. After the curing and air drying procedure the concrete samples were exposed to high temperatures 200 %C, 400 %C, 600 %C or 800 %C in an electric furnace and then cooled to the room temperature. This was followed by experimental investigation using non-destructive and destructive test methods. The reference values of the experimental measurements were determined on a non-preheated group of test specimens. The results of the non-destructive tests include the determination of the ultrasound (US) pulse velocity, the surface strength, the dynamic modulus of elasticity and the shear modulus of concrete, while destructive tests include the determination of the compressive and flexural strengths and the modulus of elasticity of concrete. Using statistical methods it was then determined that temperature has a statistically significant influence on the above mentioned experimental results, meaning that changes between individual preheated specimens can be detected. This was followed by an estimation of the mechanical properties of concrete after exposure to high temperatures, also named residual mechanical properties, using regression models with explicit relationships and artificial neural networks. We found that the residual flexural strength and modulus of elasticity of concrete can be estimated very accurately based on regression models with explicit relationships, whereas a more accurate estimation of residual compressive strength requires the use of artificial neural networks. In the second part a numerical model for the determination of the fire resistance of planar RC structures after a fire, named Nfira, is briefly presented. The novelty of the numerical model are the experimentally determined material parameters of the constitutive law of limestone concrete after exposure to high temperatures. This was followed by the parametric studies in which the influence of different fire scenarios and concrete mixture on the behavior of planar RC structures after fire were investigated.In our dissertation we dealt with the estimation of the load bearing capacity of reinforced concrete (RC) frame structures after fire, which we carried out in two parts. In the first part we presented various non-destructive and destructive methods for concrete testing used in the experimental part. Within the experimental investigation we prepared five different concrete mixtures with limestone aggregate, which differ in water to cement ratio, type of cement and the amount of water and additives. After the curing and air drying procedure the concrete samples were exposed to high temperatures 200 %C, 400 %C, 600 %C or 800 %C in an electric furnace and then cooled to the room temperature. This was followed by experimental investigation using non-destructive and destructive test methods. The reference values of the experimental measurements were determined on a non-preheated group of test specimens. The results of the non-destructive tests include the determination of the ultrasound (US) pulse velocity, the surface strength, the dynamic modulus of elasticity and the shear modulus of concrete, while destructive tests include the determination of the compressive and flexural strengths and the modulus of elasticity of concrete. Using statistical methods it was then determined that temperature has a statistically significant influence on the above mentioned experimental results, meaning that changes between individual preheated specimens can be detected. This was followed by an estimation of the mechanical properties of concrete after exposure to high temperatures, also named residual mechanical properties, using regression models with explicit relationships and artificial neural networks. We found that the residual flexural strength and modulus of elasticity of concrete can be estimated very accurately based on regression models with explicit relationships, whereas a more accurate estimation of residual compressive strength requires the use of artificial neural networks. In the second part a numerical model for the determination of the fire resistance of planar RC structures after a fire, named Nfira, is briefly presented. The novelty of the numerical model are the experimentally determined material parameters of the constitutive law of limestone concrete after exposure to high temperatures. This was followed by the parametric studies in which the influence of different fire scenarios and concrete mixture on the behavior of planar RC structures after fire were investigated.

In this paper the possibility of using different regression models to predict the mechanical properties of limestone concrete after exposure to high temperatures, based on the results of non-destructive techniques, that could be easily used in-situ, is discussed. Extensive experimental work was carried out on limestone concrete mixtures, that differed in the water to cement (w/c) ratio, the type of cement and the quantity of superplasticizer added. After standard curing, the specimens were exposed to various high temperature levels, i.e., 200%, 400%, 600% or 800%. Before heating, the reference mechanical properties of the concrete were determined at ambient temperature. After the heating process, the specimens were cooled naturally to ambient temperature and tested using non-destructive techniques. Among the mechanical properties of the specimens after heating, known also as the residual mechanical properties, the residual modulus of elasticity, compressive and flexural strengths were determined. The results show that residual modulus of elasticity, compressive and flexural strengths can be reliably predicted using an artificial neural network approach based on ultrasonic pulse velocity, residual surface strength, some mixture parameters and maximal temperature reached in concrete during heating

Strojno učenje se vse pogosteje pojavlja kot rešitev problemov, ki jih je težko rešiti s klasičnimi pristopi. Pri strojnem učenju algoritem naučimo na osnovi podatkov, namesto da bi ga eksplicitno napisali. V tej nalogi smo z metodami strojnega učenja razvili krmilnik mobilnega robota za sledenje črti. Razvili smo zaznavalo črte, ki deluje odlično v različnih pogojih. Ustvarili smo simulacijsko okolje za preizkušanje krmiljenja robota in v njem z metodami vzpodbujevalnega učenja razvili krmilnik, ki sledi črti. Krmilnik smo iz simulacije prenesli v realni svet. V simulaciji krmilnik deluje primerljivo s klasičnim PID krmilnikom, v realnem svetu pa precej slabše, kar bi lahko izboljšali v nadaljnjem delu. Zaznavalo črte in krmilnik sta sposobna realnočasovnega delovanja na omejenih platformah, kot je Raspberry Pi. Machine learning is increasingly emerging as a solution to problems that are difficult to solve with classical approaches. In machine learning, the algorithm is trained on data, rather than written explicitly. In this thesis, we developed a mobile robot controller using methods of machine learning. We developed a line detector, that works well under different conditions. We created a simulation environment, designed for testing robot control algorithms, and developed a line following controller using reinforcement learning methods. The controller was transferred from simulation to the real world. In the simulation, controller performance is comparable to the classic PID controller, whereas in the real world, it is considerably worse. That could be improved in future work. The detector and controller are capable of real-time operation on limited platforms such as Raspberry Pi.

Z razvojem področja globokega učenja, ki temelji na umetnih nevronskih mrežah, se danes poskušajo rešiti že znani problemi področja obdelave naravnega jezika. V tem magistrskem delu obravnavamo problem razpoznavanja in klasifikacije imenskih entitet z uporabo metod globokega učenja. V magistrski nalogi smo uporabili programski jezik Python in odprtokodno knjižnico Keras. Preizkusili smo različne arhitekture rekurentnih nevronskih mrež, ki uporabljajo pomnilne celice LSTM in GRU. Prav tako smo opravili različne poskuse, v katerih smo iskali optimalne parametre nevronske mreže z namenom natančnega razpoznavanja in klasifikacije imenskih entitet. Učenje nevronske mreže in vrednotenje modelov smo izvedli na korpusih, ki so bili predstavljeni na konferenci CONLL leta 2003. Deep learning growth based on artificial neural networks allowed us to solve well-known problems in the natural language processing field. In this Master's thesis we deal with the problem of identifying and classifying named entities using deep learning methods. In the project, we used the Python programming language and the Keras library. We tested different architectures of recurrent neural networks that use LSTM and GRU memory cells. We also performed various experiments in which we searched for the optimal parameters of the neural network with the intent to accurately recognize and classify name entities. Neural network learning and model evaluation were conducted at the corpora presented at the CONLL conference in 2003.

Hiter razvoj na področju vgrajenih sistemov v zadnjih letih danes omogoča vgrajevanje zmogljivih procesnih enot v mnoge naprave. Priprava hrane je eno od področij, kjer lahko uporaba takšnih tehnologij olajša vsakodnevna opravila ter prispeva k boljšim rezultatom. V diplomski nalogi smo testirali možnost uporabe kamere in umetnih nevronskih mrež za nadzor stepanja smetane v naprednih kuhinjskih mešalnikih. Najprej smo naredili več posnetkov stepanja smetane ter posamezne fotografije iz posnetkov označili glede na stopnjo stepenosti smetane. Predstavili smo nekaj pomembnejših nevronskih mrež za analizo fotografij (AlexNet, MobileNet, NasNet) ter testirali, kako se nevronske mreže, naučene na podlagi ustvarjene baze fotografij, obnesejo v praksi. Rezultati so pokazali, da lahko tako naučena nevronska mreža doseže primerno točnost, tudi ko so posnetki narejeni v drugačnih pogojih kot posnetki, ki smo jih uporabili pri učenju. Rapid development of embedded systems in recent years allows us to integrate powerful processing units in most electronic devices. Food preparation is one such field, where technology can ease everyday chores, and help us achieve better results. We've tried out various ways to use a camera in combination with artificial neural networks to control kitchen mixer when making whipped cream. First we made several recordings of cream during mixing. We've labelled each frame according to how well-mixed the cream is. We show some higher-importance neural networks for image analysis (AlexNet, MobileNet, NasNet) and test how well those neural networks perform after being trained on our dataset. Our results indicate that such neural networks are able to give an accurate prediction, even on photos captured under different conditions than the training data.

Za analizo povezav med lesnimi branikami in okoljem so primerjane (multipla) linearna regresija (MLR) in izbrane nelinearne metode s področja strojnega učenja: umetne nevronske mreže z učnim algoritmom Bayesova regularizacija (ANN), modelna drevesa (MT), ansambli modelnih dreves (BMT) in naključni gozdovi regresijskih dreves (RF). Izbrane metode so bile primerjane na devetih podatkovnih množicah, ki so vključevale različne parametre branik in različne okoljske spremenljivke. Za nelinearne metode so bili v večini primerov izračunani boljši statistični kazalci za validacijsko množico, čeprav razlike v primerjavi z linearno regresijo niso bile velike. Dodatne analize so pokazale, da se metode večinoma razlikujejo pri napovedih ekstremnih vrednosti. Značilnost nelinearnih metod je, da sprememba odvisne spremenljivke ni premo sorazmerna spremembi ene ali več neodvisnih spremenljivk. Slednje pomeni zmanjšan razpon in variabilnost rekonstruiranih vrednosti, kar rekonstrukcijo naredi vizualno manj privlačno od linearne ekstrapolacije, čeprav v večini primerov statistično boljšo. Nobena izmed nelinearnih metod s področja strojnega učenja ni dala najboljših rezultatov na vseh podatkovnih množicah, zato je pred rekonstrukcijo klime z metodami strojnega učenja vedno smiselno primerjati več metod. Za ta namen smo razvili R funkcijo compare_methods() in jo vgradili v dendroTools R paket, ki je prosto dostopen na CRAN-repozitoriju. (Multiple) linear regression (MLR) and selected nonlinear methods from the field of machine learning were compared for the analysis of relationships between xylem tree-rings and the environment: artificial neural networks with a training algorithm that uses Bayesian regularization (ANN), model trees (MT), ensembles of model trees (BMT) and random forests of regression trees (RF). The selected methods were compared on nine datasets, which included different tree-ring parameters and different target environmental variables. For the nonlinear methods, better statistical metrics were calculated on validation data in most cases, but the differences in comparison to linear regression were minor. Additional analysis indicated that the methods mostly differ in predicting the extreme values. The characteristic of nonlinear methods is that the change in the dependent variable is not proportional to the change of one or more independent variables. The latter results in a reduced range and variability of reconstructed values, which makes the reconstruction visually less attractive as compared to linear extrapolation, even though in most cases statistically better. None of the nonlinear machine learning methods showed best results on all datasets, therefore it makes sense to always compare different machine learning regression methods prior to climate reconstruction. To do so, the R function compare_methods() was developed and implemented in the dendroTools R package, which is freely available on the CRAN repository.

Na področju robotike obstaja ogromno nelinearnih sistemov, ki se še vedno vodijo z linearnimi regulatorji, čeprav ti niso optimalna rešitev za dani problem. V tem magistrskem delu je predstavljen hitrostni adaptivni nelinearni regulator, ki je sposoben voditi nelinearno progo boljše kot linearni regulatorji. Razviti regulator je sestavljen iz algoritma po vzorih iz narave, ki optimira vrednost referenčnega toka, in umetne nevronske mreže, ki je sposobna napovedati vrednost ocenitvene funkcije za izbrani algoritem. Pri tem primerjamo vpliv različnih algoritmov po vzorih iz narave na delovanje predlaganega regulatorja. V naši primerjalni analizi smo zajeli naslednje algoritme: evolucijsko strategijo, diferencialno evolucijo, optimizacijo z roji delcev in algoritmom po vzoru obnašanja netopirjev. In robotics, there are a lot of nonlinear systems, which are still controlled using linear controllers, even though they are not optimal solutions for the given problem. In this work, an adaptive nonlinear velocity controller is presented, which is better suited for control of nonlinear systems. The presented controller consists from nature inspired algorithm, which optimizes current reference, and artificial neural network, which is used for fitness function evaluation. A comparison of controller operation, when different nature inspired algorithms are used, is also presented in this work. Thus, the following nature inspired algorithms were captured in our comparison study: evolutionary strategy, differential evolution, particle swarm algorithm and bat algorithm.

Cilj podjetij in investitorjev na finančnih trgih je bolj ali manj enak, doseganje čim večjih dobičkov. To je glavni razlog za razvoj metod, ki bi čim bolj natančno napovedale donose investiranih sredstev. Sprva je gibanje cen vrednostnih papirjev veljalo za povsem nepredvidljivo, kar je vodilo do razvoja teorije učinkovitega trga kapitala. Tekom let pa so empirične raziskave na finančnem področju pokazale, da na trgu obstajajo določene anomalije. S pomočjo analize gibanja cen v preteklosti si investitor lahko zastavi trgovalno strategijo, ki mu bo na dolgi rok prinašala nadpovprečne donose v primerjavi s trgom. Obstaja veliko metod za analizo in napovedovanje, a z razvojem računalništva v ospredje stopa strojno učenje. Ena izmed najbolj razširjenih metod strojnega učenja so umetne nevronske mreže. S pomočjo strojnega učenja je mogoče natančneje modelirati anomalije na trgu kapitala in iskati morebitne povezanosti med cenami vrednostnih papirjev, kot pa z uporabo tradicionalnih metod. V magistrskem delu je predstavljeno teoretično ozadje predvidljivosti v gibanju cen delnic in metode napovedovanja gibanja cen delnic. Poudarek je na integriranem avtoregresijskem modelu s premikajočim povprečjem (ARIMA model) in metodi nevronskih mrež. ARIMA model je kombinacija linearnih modelov časovnih vrst, in sicer avtoregresijskega modela (AR model) in modela premikajočega povprečja (MA model). Nevronske mreže pa so inteligentni sistemi, ki posnemajo delovanje živčnih celic v možganih. Zgrajene so iz umetnih nevronov in njihova glavna lastnost je sposobnost učenja povezave med vhodnimi in izhodnimi podatki. Omenjeni metodi sta uporabljeni na dejanskih podatkih ameriškega trga, in sicer na podatkih indeksa Dow Jones Industrial Average (DJIA) iz obdobja 2009 - 2014. Indeks DJIA kotira na newyorški borzi in na borzi NASDAQ ter zajema trideset najpomembnejših delnic newyorške borze. S pomočjo metod smo napovedali gibanje vrednosti indeksa ob koncu dneva v letu 2014. Ob primerjavi dobljenih rezultatov metod smo ugotovili, da metoda nevronskih mrež da veliko boljše rezultate napovedi kot ARIMA model. Na podlagi analize dobljenih rezultatov bi investitorju predlagali, da si pri napovedovanju gibanja cen delnic in s tem pri izbiranju najboljše strategije trgovanja z delnicami, pomaga z metodo nevronskih mrež. The goal of companies and investors in financial markets is more or less the same, achieving the highest possible profits. This is the main reason for the development of methods, which would predict returns of invested assets as accurately as possible. Initially, the movement of securities prices was considered completely unpredictable, which led to the development of efficient capital market theory. During the years, empirical researches in the financial field has shown, that the capital market does not meet all assumptions of efficient capital market and that there exist certain anomalies. With the help of the analysis of past price movement, investor can set a trading strategy that will bring him above-average returns compared to the market in the long run. Many methods have been developed for analyzing the movement of securities prices and forecasting prices in the future. With the development of computer science, machine learning is becoming more and more popular. One of the most commonly used methods of machine learning are artificial neural networks. Machine learning enables more precise modeling of anomalies in the capital market and easier definition of possible connections between securities prices, compared to the traditional methods. In the master thesis the theoretical background of predictability in the movement of share prices and methods for forecasting their movement in the future are presented. The focus is on Autoregressive integrated moving average model (ARIMA model) and artificial neural networks. ARIMA model is a combination of time series linear models, namely autoregressive model (AR model) and moving average model (MA model). Artificial neural networks are intelligent systems that imitate the functions of nerve cells in the human brains. They are built from artificial neurons and their main characteristic is the ability to learn the connection between input and output data. Both mentioned methods are applied on the actual USA market data, namely on Dow Jones Industrial Average (DJIA) data from the period 2009-2014. DJIA index trade on the New York Stock Exchange (NYSE) and on NASDAQ. It consists of thirty major shares traded on the NYSE. We made the forecast of the DJIA index value movement in the year 2014. The comparison of the obtained results has shown, that artificial neural networks provide better results than ARIMA model. Based on our findings, we would propose an investor to choose artificial neural networks for share prices movement forecasting. Consequently, the investor could choose the best trading strategy.

Različne študije so pokazale, da lahko z nelinearnimi metodami bolje opišemo (modeliramo) odnos med branikami in okoljem. V naši študiji smo primerjali (multiplo) linearno regresijo (MLR) in štiri nelinearne metode strojnega učenja: modelna drevesa (MT), ansambel bagging modelnih dreves (BMT), umetne nevronske mreže (ANN) in metodo naključnih gozdov (RF). Za primerjavo teh metod modeliranja smo uporabili štiri množice podatkov. Natančnost naučenih modelov smo ocenili z metodo 10-kratnega prečnega preverjanja (ang. 10-fold cross-validation) na naši množici in preverjanjem na dodatni testni množici. Na vseh množicah smo dobili boljše statistične kazalce za nelinearne metode s področja strojnega učenja, s katerimi lahko pojasnimo večji delež variance oz. dobimo manjšo napako. Nobena metoda se ni pokazala kot najboljša v vseh primerih, zato je smiselno predhodno primerjati več različnih metod in nato uporabiti najprimernejšo, npr. za rekonstrukcijo klime. Many studies have shown that by using nonlinear methods, the relationship between tree-ring parameters and the environment can be described (modelled) better and in more detail. In our study, (multiple) linear regression (MLR) with four nonlinear machine learning methods are compared: artificial neural networks (ANN), model trees (MT), bagging of model trees (BMT) and random forests of regression trees (RF). To compare the different regression methods, four datasets were used. The performance of the learned models was estimated by using 10-fold cross-validation and an additional hold-out test. For all datasets, better results were obtained by the nonlinear machine learning regression methods, which can explain more variance and yield lower error. However, none of the considered methods outperformed all other methods for all datasets. Therefore, we suggest testing several different methods before selecting the best one, e.g. for climate reconstruction.

This paper describes a computer vision system based on image processing and machine learning techniques which was implemented for automatic assessment of the tomato seed germination rate. The entire system was built using open source applications Image J, Weka and their public Java classes and linked by our specially developed code. After object detection, we applied artificial neural networks (ANN), which was able to correctly classify 95.44% of germinated seeds of tomato (Solanum lycopersicum L.). Članek opisuje sistem računalniškega vida, ki temelji na tehnikah obdelave slik in strojnega učenja, ki je bil izdelan za avtomatsko oceno stopnje kaljenja semen paradižnika. Celoten sistem je bil zgrajen s pomočjo odprtokodnih aplikacij ImageJ, Weka in njihovih javno dostopnih javanskih kod, ki smo jih povezali v lastno originalno razvito kodo. Po odkrivanju predmetov na RGB slikah, smo uporabili umetne nevronske mreže (ANN), ki so bile sposobne pravilno razvrstiti 95,44% nakaljenih semen paradižnika (Solanum lycopersicum L.).

Odpadne vode iz tekstilne industrije, ki nastajajo pri plemenitenju tekstilnih materialov, so zaradi svoje obarvanosti in heterogene sestave tako estetski kot tudi ekološki problem. V članku je opisan evropski projekt ADOPBIO, ki ponuja alternativno rešitev čiščenja teh odpadnih voda. Namen projekta ADOPBIO je razviti in ponuditi rešitev čiščenja in recikliranja odpadne vode iz tekstilne plemenitilne industrije. Cilj projekta je razviti metodo, ki zagotavlja popolno razbarvanje in 75-odstotno recikliranje odpadne vode iz plemenitilnic. ADOPBIO združuje dve najboljši razpoložljivi tehniki za čiščenje odpadne vode: tehniko AOP z bioflotacijo. Ta kombinacija metod čiščenja odpadnih voda iz tekstilne industrije do začetka projekta še ni bila preizkušena. Vloga Oddelka za tekstilne materiale in oblikovanje pri tem projektu je bila raziskati in proučiti možnost razbarvanja odpadnih voda iz dveh sodelujočih tekstilnih podjetij z naprednima oksidacijskima postopkoma: s postopkom, pri katerem aktiviramo $H_2O_2$ termično ob prisotnosti katalizatorja, in s postopkom, kjer aktiviramo $H_2O_2$ z UV žarki. Z obdelavo termo/$H_2O_2$/katalizator, ki ji je sledila obdelava UV/$H_2O_2$, smo za odpadne vode iz podjetja TSP v povprečju dosegli 70-odstotno razbarvanje, za odpadne vode iz podjetja Blondel pa 87-odstotno razbarvanje. Due to its intense coloration and heterogeneous composition, textile finishing wastewaters represent an aesthetic and an ecological problem. In this paper the European project ADOPBIO, which offers an alternative solution for textile wastewater treatment, is presented. The objective of ADOPBIO project is to develop a solution for treatment and recycling of textile finishing wastewater. The aim of the project is to develop a method that would enable total decoloration and 75% wastewater recycling. The ADOPBIO project comprises two optimal treatment techniques: AOP and bioflotation. The combination of both techniques has not been tested as yet. The role of the Department of Textile Materials and Design (University of Maribor) in this project was to investigate the possibilities of textile wastewater decoloration by using thermally activated $H_2O_2$ in presence of catalyst and UV activated $H_2O_2$. Textile wastewater used in the study was obtained from two companies: TSP and Blondel. By using the thermal/$H_2O_2$/catalyst process followed by the UV/$H_2O_2$ process a 70% decoloration of wastewater from TSP and 87% decoloration of Blondel wastewater was achieved.

Chemical composition of Slovenian coal has been characterised in terms of proximate and ultimate analyses and the relations among the chemical descriptors and the higher heating value (HHV) examined using correlation analysis and multivariate data analysis methods. The proximate analysis descriptors were used to predict HHV using multiple linear regression (MLR) and artificial neural network (ANN) methods. An attempt has been made to select the model with the optimal number of predictor variables. According to the adjusted multiple coefficient of determination in the MLR model, and alternatively, according to sensitivity analysis in ANN developing, two descriptors were evaluated by both methods as optimal predictors: fixed carbonand volatile matter. The performances of MLR and ANN when modelling HHV were comparable the mean relative difference between the actual and calculated HHV values in the training data was 1.11% for MLR and 0.91% for ANN. The predictive ability of the models was evaluated by an external validation data set the mean relative difference between the actual and predicted HHV values was 1.39% in MLR and 1.47% in ANN. Thus, the developed models could be appropriately used to calculate HHV.

The determination of the sensorial quality of wines is of great interest for wine consumers and producers since it declares the quality in most of the cases. The sensorial assays carried out by a group of experts are time-consuming and expensive especially when dealing with large batches of wines. Therefore, an attempt was made to assess the possibility of estimating the wine sensorial quality with using routinely measured chemical descriptors as predictors. For this purpose, 131 Slovenian red wine samples of different varieties and years of production were analysed and correlation and principal component analysis were applied to find inter-relations between the studied oenological descriptors. The method of artificial neural networks (ANNs) was utilised as the prediction tool for estimating overall sensorial quality of red wines. Each model was rigorously validated and sensitivity analysis was applied as a method for selecting the most important predictors. Consequently, acceptable results were obtained, when data representing only one year of production were included in the analysis. In this case, the coefficient of determination (R2) associated with training data was 0.95 and that for validation data was 0.90. When estimating sensorial quality in categorical form, 94 % and 85 % of correctly classified samples were achieved for training and validation subset, respectively.

Z opazovanjem absorpcijskega spektra, ki ga dobivamo z obsevanjem določenega vzorca, na primer tkiva, z infrardečo svetlobo, lahko dobimo informacijo o kemijski sestavi vzorca. Pri analizi spektrov pogosto uporabljamo klasifikacijske metode, s katerimi lahko določamo sestavo celotnega vzorca ali njegovih delov. Eden od primernih algoritmov za ta namen so umetne nevronske mreže. V diplomskem delu smo ugotavljali, kako uspešne so umetne nevronske mreže za klasifikacijo infrardečih spektrov. Preizkusili smo jih na več naborih podatkov ter jih primerjali z metodo podpornih vektorjev in multinomsko logistično regresijo. Preverjali smo tudi uspešnost konvolucijskih nevronskih mrež. Umetne nevronske mreže so dosegle primerno točnost, vendar niso veliko boljše od metode podpornih vektorjev, ki je, po drugi strani, bistveno hitrejša. Absorption spectra obtained from the sample irradiated by infrared radiation represent a very useful method for observing the chemical composition of different kinds of samples, from cell tissue to various materials. For spectrum analysis, we often use classification algorithms. A suitable algorithm for this task is an artificial neural network. In the diploma thesis, we explored the usefulness of artificial neural networks for classification of infrared spectra. We measured classification accuracies on different data sets and compared them to the results of support vector machines and multinomial logistic regression. We also examined the performance of convolutional neural networks. The results achieved by the artificial neural networks were promising. However, they were not significantly better than those of the support vector machines. On the other hand, the performance of the latter was considerably faster.

Digitalno trgovanje z dobrinami (delnicami, valutami itd.) danes močno izpodriva klasično trgovanje, saj se je veliko borz preselilo v oblak. Računalnik poskrbi za izmenjavo dobrin, hkrati pa lahko tudi samostojno trguje. Avtomatski trgovalni sistem je računalniški sistem, ki trguje z dobrinami brez posredovanja človeka. Tak pristop poleg objektivnosti in empiričnosti odločitev omogoča hitre izvršitve kupčij, ki so ključne za uspeh. V diplomski nalogi raziskujemo avtomatske trgovalne sisteme in njihov način delovanja. Naredimo pregled področja tehnične analize trgovanja, ki se ukvarja s kvantificiranjem gibanja cen na trgu. Definiramo problem trgovanja z vidika nadzorovanega strojnega učenja in rudarjenja podatkovnih tokov. S področja strojnega učenja posebej izpostavimo algoritem k-najbližjih sosedov, umetne nevronske mreže in naivni Bayesov klasifikator. S pomočjo omenjenih algoritmov in podatkov, pridobljenih z metodami tehnične analize trgovanja, zasnujemo avtomatski trgovalni sistem. S simulacijo ga ovrednotimo na realnem gibanju cen kriptovalute Bitcoin in kriptovalute Litecoin. Avtomatsko trgovanje se z vidika strojnega učenja izkaže za zelo zahteven problem. Kljub temu nam uspe, z uporabo algoritma k-najbližjih sosedov in umetnih nevronskih mrež, doseči zadovoljivo napovedno uspešnost in posledično profitabilnost. Digital trading of securities is beginning to dominate over classical trading and the trading exchanges are rapidly migrating to the cloud. Computer is not only present in the exchange process but is also capable of making trading decisions on human's behalf. Automated trading system is a computer system, capable of trading without human interaction. The benefits of such an approach to trading are objectivity and fast execution of orders, which are often crucial for success. In this thesis we examine automated trading systems and their structure. We study the field of technical analysis which quantifies market price movements. We define trading as a supervised machine learning and stream mining problem and examine the k-nearest neighbours algorithm, naive Bayes classifier and artificial neural networks. Based on our research we design an automated trading system. We evaluate its performance on actual market data of cryptocurrencies Bitcoin and Litecoin using a simulated environment. Automated trading turns out to be a difficult machine learning problem, but with the use of the k-nearest neighbours algorithm and artificial neural networks we manage to achieve decent success in our predictions and profitability.

SI: Navigacijski sistemi običajno temeljijo na sprejemnikih GNSS. Članek predstavlja trenutno stanje sistemov GNSS, navigacijo s tehnologijami GNSS in opisuje GPS-navigacijske instrumente. Nekatere naloge zahtevajo neprekinjeno navigacijo, česar ne moremo zagotoviti samo z uporabo GNSS-navigacije. Sistem za neprekinjeno navigacijo dopolnjujejo inercialni navigacijski sistemi. Predstavljene so osnove inercialne navigacije, opisane so inercialne merilne enote,navedeni tipi IMU in tipični pogreški inercialnih senzorjev. Predstavljeni sta dve metodi obdelave podatkov združenih sistemov GNSS/INS, tradicionalni Kalmanov filter in umetne nevronske mreže, ki po nekaterih raziskavah dosegajo boljše rezultate kot Kalmanov filter. Omenjeni sta še dve dodatni možnosti za izboljšanje ali celo zagotovitev neprekinjene navigacije, psevdoliti in širokopasovni radijski valovi. EN: Navigation systems are commonly based on GNSS receivers. The article presents the current status of GNSS, discusses GNSS navigation and describes GPS navigation instruments. Some applications require seamless navigation which cannot be provided byGNSS itself. The system for uninterrupted navigation uses Inertial Navigation System besides GNSS. The bases of inertial navigation, Inertial Measurement Units, types of IMU and typical inertial sensorerrors are presented. The processing of data from an integrated GNSS/INS is usually performed either by a traditional Kalman Filter or an artificial neural network. According to some of the research, the latter performs better than the conventional Kalman Filter.Seamless navigation can be improved or even made possible by the use of pseudolites or Ultra-wide Band.

U radu se predstavlja razvijeni inteligentni sustav za predviđanje mehaničkih svojstava materijala na temelju metalografskih slika. Sustav se sastoji od dva modula. Prvi je modul algoritam za dobivanje karakteristika iz metalografskih slika. Prvi algoritam očitava metalografsku sliku dobivenu mikroskopom, zatim se dobivaju karakterisike razvijenim algoritmom, i na kraju algoritam izračunava omjere mikrostrukture materijala. U ovom istraživanju potrebno je što točnije odrediti omjere grafita, ferita i ausferita iz metalografskih slika. Drugi modul razvijenog sustava je sustav za predviđanje mehaničkih svojstava materijala. Predviđanje mehaničkih svojstava materijala izvršeno je pomoću feed-forward umjetne neuronske mreže. Kao ulazi u umjetnu neuronsku mrežu rabljeni su izračunati omjeri grafita, ferita i ausferita, dok su mehanička svojstva materijala upotrebljena kao ciljevi za uvježbavanje. Uvježbavanje umjetnih neuronskih mreža obavljeno je na prilično maloj bazi podataka, no mijenjajući parametre nama je to uspjelo. Umjetna neuronska mreža je naučila do te mjere da je greška bila prihvatljiva. S orijentiranom neuronskom mrežom uspješno smo predvidjeli mehanička svojstva izuzetog uzorka., This article presents developed intelligent system for prediction of mechanical properties of material based on metallographic images. The system is composed of two modules. The first module of the system is an algorithm for features extraction from metallographic images. The first algorithm reads metallographic image, which was obtained by microscope, followed by image features extraction with developed algorithm and in the end algorithm calculates proportions of the material microstructure. In this research we need to determine proportions of graphite, ferrite and ausferrite from metallographic images as accurately as possible. The second module of the developed system is a system for prediction of mechanical properties of material. Prediction of mechanical properties of material was performed by feed-forward artificial neural network. As inputs into artificial neural network calculated proportions of graphite, ferrite and ausferrite were used, as targets for training mechanical properties of material were used. Training of artificial neural network was performed on quite small database, but with parameters changing we succeeded. Artificial neural network learned to such extent that the error was acceptable. With the oriented neural network we successfully predicted mechanical properties for excluded sample.

magistrsko delo

Determination of the product's origin is one of the primary requirements when certifying a wine's authenticity. Significant research has described the possibilities of predicting a wine's origin using efficient methods of wine components' analyses connected with multivariate data analysis. The main goal of this study was to examine the discrimination ability of simple enological descriptors for the classification of Slovenian red and white wine samples according to their varieties and geographical origins. Another task was to investigate the inter-relations available among descriptors such as relative density, content of total acids, non-volatile acids and volatile acids, ash, reducing sugars, sugar-free extract, $SO_2$, ethanol, pH, and an important additional variable - the sensorial quality of the wine, using correlation analysis, principal component analysis (PCA), and cluster analysis (CLU). 739 red and white wine samples were scanned on a Wine Scan FT 120, from wave numbers 926 $cm^{–1}$ to 5012 $cm^{–1}$. The applied methods of linear discriminant analysis (LDA), general discriminant analysis (GDA), and artificial neural networks (ANN), demonstrated their power for authentication purposes. Določevanje izvora je ena od osnovnih zahtev, ko želimo certificirati pristnost vin. Raziskava opisuje možnosti napovedovanja izvora vin z uporabo učinkovitih metod analize parametrov vin in multivariantno analizo. Glavni namen študije je proučevanje možnosti razlikovanja enostavnih enoloških deskriptorjev za klasifikacijo vzorcev slovenskih rdečih in belih vin glede na vrsto in geografski izvor. Drugi cilj je bil proučevanje razmerij med deskriptorji, kot so: relativna gostota, vsebnost skupnih kislin, nehlapne kisline, hlapne kisline, pepel, reducirajoči sladkor, prosti sladkor, $SO_2$, etanol, pH in med pomembnimi dodatnimi spremenljivkami, kot je senzorična kakovost vina z uporabo korelacijske analize, metode glavnih osi (PCA) in analizo grupiranja podatkov (CLU). 739 vzorcev rdečih in belih vin je bilo posnetih na aparatu Wine Scan FT 120, od valovnega števila 926 $cm^{–1}$ do 5012 $cm^{–1}$. Uporabljene metode linearne diskriminantne analize (LDA), splošne diskriminantne analize (GDA) in umetnih nevronskih mrež (ANN) potrjujejo sposobnost določanja pristnosti vin.

V magistrskem delu smo uporabili algoritme po vzorih iz narave za finančna modeliranja. Najprej smo uporabili umetne nevronske mreže za napovedovanje cene delnice, nato pa še genetske algoritme za optimizacijo portfelja delnic, ki smo jih primerjali s kvadratnim programiranjem. V raziskavi se je izkazalo, da lahko s umetnimi nevronskimi mrežami bolje ocenimo variančno-kovariančno matriko, kot če bi uporabili zgodovinske podatke. Pri reševanju problema optimizacije portfelja delnic se je izkazalo, da lahko z genetski algoritmi dobimo rezultate primerljive s kvadratnim programiranjem, saj rezultati med tehnikama, predvsem pri manjšem porteflju, v glavnem niso statistično značilni. In a master work we used nature inspired algorithms for financial modeling. Firstly we use an artificial neural networks to predict stock prices and secondly, we used genetic algorithms for stock portfolio optmization. The results show better assessing covariance matrix with neural networks gives more accurate results in a portfolio optimization than if we are taking historical prices. We can assert that results obtained with a genetic algorithms are in general statistically the same as they are with quadratic programming, especially in cases with less stocks in a portfolio.

Namen magistrskega dela je bil izdelati model lokalne sekundarne regulacije napetosti na osnovi umetnih nevronskih mrež (LSRN-UNM) in sposobnost regulacije modela preveriti s simulacijo na modelu elektroenergetskega sistema (EES), ki ima spremenljivo količino inštalirane moči sončnih elektrarn (SE). V magistrski nalogi smo najprej predstavili osnove delovanja regulacije napetosti v EES. Nato smo opisali metode za analizo modela EES in pri tem tudi definirali testni model IEEE RTS. Sledil je opis strukture UNM ter predstavitev Levenberg-Marquardtovega algoritma, ki smo ga uporabili za učenje. Učno množico, ki smo jo dobili z velikim številom simulacij modela EES smo uporabili za učenje LSRN-UNM, kateremu je sledila integracija naučenih LSRN-UNM v model EES. Opazovali smo delovanje LSRN-UNM v različnih letnih časih in za različne deleže sončnih elektrarn v EES. Ugotovili smo dobro delovanje LSRN-UNM, razen v primerih, ko je bil EES v stanju, ki ni bilo zajeto v učni množici. Objective of this master thesis was to design a model of local secondary voltage control with artificial neural networks and to verify its voltage control capability with simulations using a power system model with variable share of solar power plants. Firstly, we described the basics of voltage control in power systems. Afterwards, methods for power systems analysis and test power system IEEE RTS were defined. Artificial neural networks and Levenberg-Marquardt learning algorithm, which was used for learning process, were also described. Training data was acquired with large number of simulations and then used for learning of artificial neural networks of local secondary voltage control, which were then integrated in power system model. We observed local secondary voltage control for different seasons and different shares of solar power plants. Local secondary voltage control operated well, except when power system model was in a state not covered by training data.

V disertaciji je predstavljena možnost uporabe umetnih nevronskih mrež za receptiranje barvil v tiskarskih barvnih goščah. V nalogi su obdelani podtaki barvnih vzorcev iz barvne karte, ki jo v tekstilnem podjetju uporabljajo za vizualno receptiranje. Vzorci so bili izmerjeni, urejeni po kombinacijah dveh barvil in predelani v obliko, primerno za obdelavo z nevronskih mrežami. V prvem delu naloge je raziskana možnost napovedanja barvnih kombinacij na podlagi refleksijskih vrednosti. 1430 vzorcev, potiskanih s kombinacijami dveh barvil izmed desetih, je bilo uporabljenih za treniranje protitočne nevronske mreže. Najboljši model je v 77% pravilno napovedal obe barvil v kombinaciji, v 22, 3% je pravilno napovedal eno barvilo in v 0, 7% obe barvil napačno. Z "boot-strap" navzkrižnim preverjanjem je bil model validiran s korelacijo na 0, 7. Opravljena je bila primerjava z barvnometričnimi metodami. Drugi delu naloge opisuje izdelavo modela za napovedanje koncentracij barvil na vzorcu. Vhodni podatki v mreže z vzvrtanim širjenjem napake so refleksije ali barvne vrednosti vzorcev iz barvne karte. Za vsako kombinacijo dveh barvil (po 46 vzorcev) je potrebna specifična izdelava modela (izbor vzorcev, določitev parametrov mreže). Prikazani so modeli za dve različni kombinaciji. Raziskan je vpliv vrste in količine podatkov ter parametrov nevronske mreže na uspešnost učenja. Preizkušene su bile razne strukture nevronskih mrež in različni parametri učenja, ugotovljene optimalne nastavitve za posamezne skupine podatkov in podane splošne ocene uspešnosti učenja. V zaključku so podane tudi ugotovitve o primerenosti uporabe nevronskih mrež za receptiranje v tiskarskih barvnih goščah in pripombe o načinu dela.