FER Repository
Not a member yet
12128 research outputs found
Sort by
Dynamic control of simulated traffic systems using genetic programming
No full textOvaj rad razmatra primjenu algoritma genetskog programiranja na problem upravljanja prometnim signalima. Pomoću genetskog programiranja dobivaju se dinamički kontroleri koji upravljaju trajanjima faza prometnih signala na pojedinim raskrižjima u stvarnom vremenu. Dinamički kontroleri uče se na prometnim scenarijima simuliranim pomoću simulatora SUMO (Simulation of Urban MObility). Rad algoritma ispituje se na nekoliko umjetno generiranih scenarija te na jednom scenariju iz stvarnog svijeta. Rezultati pokazuju da je algoritam vrlo prilagodljiv i učinkovit. Najbolja dobivena rješenja ostvaruju od 13-30% manje kašnjenje u odnosu na klasični dinamički pristup. Međutim, nedovoljna raznolikost prometa u scenarijima za treniranje često dovodi do smanjene generalizacije. Za dobivanje potpuno prilagodljivih kontrolera potrebno je uvođenje kvalitetnijih scenarija za treniranje.This thesis aims to use genetic programming in order to solve the traffic signal control problem. Genetic programming is used to generate dynamic controllers, which control the traffic light phase lengths of specific intersections in real time. Dynamic controllers are learned by applying the algorithm on traffic scenarios simulated with SUMO (Simulation of Urban MObility). The performance of the algorithm is tested on several artificially generated scenarios, as well as one real life scenario. The results show that the algorithm is quite adaptable and effective. The best solutions perform 13-30% better than classical dynamic solutions. However, the lack of diverse traffic in training scenarios leads to worse generalization. Learning fully adaptable controllers requires higher-quality training scenarios
Machine Learning-Based Estimation of Solar Power Plant Production and Household Consumption
No full textOvaj rad analizira odnos između proizvodnje električne energije iz solarne elektrane i potrošnje kućanstva, s ciljem optimizacije samopotrošnje i smanjenja ovisnosti o elektroenergetskoj mreži. Kroz temeljitu pripremu i čišćenje povijesnih podataka, razvijen je regresijski model temeljen na polinomnoj regresiji, pri čemu su izostavljeni meteorološki podaci kako bi se naglasili unutarnji obrasci energetskog sustava. U radu su uspoređene različite metode modeliranja, uključujući polinomnu regresiju različitih stupnjeva i pomični prosjek, a posebna pažnja posvećena je numeričkoj stabilnosti rješenja korištenjem QR dekompozicije. Rezultati pokazuju izražene sezonske trendove u proizvodnji, dok potrošnja ostaje neredovita i slabije korelira s proizvodnjom, što potvrđuje potrebu za dodatnim strategijama optimizacije i mogućim ulaganjima u sustave za pohranu energije. Rad ističe važnost kvalitetne pripreme podataka i primjene stabilnih numeričkih metoda, ali i ograničenja modela bez vanjskih varijabli, te daje preporuke za budući razvoj uključivanjem meteoroloških i tržišnih podataka.This paper analyzes the relationship between electricity production from a solar power plant and household consumption, with the aim of optimizing self-consumption and reducing dependence on the power grid. Through thorough preparation and cleaning of historical data, a regression model based on polynomial regression was developed, with meteorological data omitted to emphasize the internal patterns of the energy system. The paper compares different modeling methods, including polynomial regression of various degrees and moving average, with special attention given to the numerical stability of the solution using QR decomposition. The results show pronounced seasonal trends in production, while consumption remains irregular and less correlated with production, confirming the need for additional optimization strategies and possible investments in energy storage systems. The paper highlights the importance of quality data preparation and the application of stable numerical methods, as well as the limitations of models without external variables, and provides recommendations for future development by including meteorological and market data
Development of software for time-series tracking of retail price changes
No full textOvaj rad istražuje izradu aplikacije za praćenje promjena cijena proizvoda kroz vrijeme, koristeći Python za dohvat podataka iz Web Archivea i ASP.NET Core MVC za
razvoj web-aplikacije. Analiziraju se tehnologije poput Bootstrap i Highcharts.js za vizualizaciju. Rad uključuje detalje implementacije, korisničke upute te funkcionalnosti
aplikacije, pružajući uvid u cijeli proces razvoja i potencijalne nadogradnjeThis paper explores the development of an application for tracking product price
changes over time, utilizing Python for data retrieval from the Web Archive and ASP.NET
Core MVC for web application development. Technologies such as Bootstrap and Highcharts.js for visualization are analyzed. The paper includes implementation details, user
instructions, and application functionalities, providing insight into the entire development process and potential upgrade
Power Series: theory and application
No full textTaylorovi redovi i redovi potencija su suvremen matematički alat koji imaju široku primjenu u inženjerstvu. U današnjici možemo koristiti te alate i u programiranju. Kroz ovaj rad je pokrivena najbitnija teorija i primjene Taylorovih redova i redova potencija. Kroz program smo vidjeli koliko su Taylorovi redovi odnosno polinomi moćan alat pri aproksimaciji funkcija.Taylor series and power series are modern mathematical tools that are widely used in engineering. Nowadays, we can use these tools in programming as well. This paper covers the most important theory and applications of Taylor series and power series. Through the program, we saw how Taylor series or polynomials are a powerful tool for approximating functions
Statistical evaluation of subjective and objective image quality assessment results
No full textOvaj rad bavi se analizom povezanosti između subjektivnih i objektivnih mjera kvalitete slike koristeći baze podataka LIVE i TID2013, koje sadrže slike s različitim vrstama degradacija i pripadajuće ljudske ocjene kvalitete. Cilj istraživanja bio je utvrditi koliko se dobro standardne objektivne mjere kao što su MSE, PSNR, SSIM, MS-SSIM i VIF, slažu sa subjektivnom procjenom kvalitete slike.
U sklopu rada provedena je statistička analiza koja je obuhvatila deskriptivne statistike, testiranje normalnosti, analizu varijanci među vrstama izobličenja, korelacijsku analizu te evaluaciju regresijskih modela za predikciju subjektivne kvalitete.
Rezultati analize pokazuju da objektivne mjere vrlo dobro prate subjektivne ocjene kada se radi o jednostavnim i globalnim izobličenjima, poput šuma i kompresije. U tim slučajevima postižu se visoke korelacije i dobra prediktivnost regresijskih modela. Nasuprot tome, kod složenijih, lokaliziranih ili nelinearnih distorzija, učinkovitost postojećih mjera značajno opada. Takvi nalazi upućuju na potrebu za primjenom naprednijih modela procjene kvalitete slike ili razvojem novih mjera osjetljivijih na kompleksne oblike degradacije.This thesis analyzes the relationship between subjective and objective image quality measures using the LIVE and TID2013 databases, which contain images with various types of degradation and corresponding human quality ratings. The aim of the study was to determine how well standard objective measures such as MSE, PSNR, SSIM, MS-SSIM, and VIF, align with subjective assessments of image quality.
A statistical analysis was conducted, covering descriptive statistics, normality testing, analysis of variance across distortion types, correlation analysis, and evaluation of regression models for predicting subjective quality.
The results show that objective measures closely follow subjective ratings in cases of simple and global distortions, such as noise and compression. In these cases, high correlation values and strong regression model performance were observed. Conversely, for more complex, localized, or nonlinear distortions, the effectiveness of current measures declines significantly. These findings highlight the need to apply more advanced models of image quality assessment or to develop new measures that are more sensitive to complex forms of degradation
Evaluating retrieval-augmented large language models in e-commerce analytics
No full textEksponencijalni rast podataka u e-trgovini i digitalnom marketingu stvorio je potrebu za pristupačnim, automatiziranim rješenjima poslovne analitike. Tradicionalni pristupi izvlačenju uvida iz strukturiranih poslovnih podataka zahtijevaju značajnu tehničku stručnost, osobito u formuliranju i izvršavanju SQL upita, što ograničava pristup donošenju odluka temeljenih na podacima. Ovaj rad istražuje primjenu otvorenih velikih jezičnih modela (LLM) kao autonomnih agenata za SQL-analitiku u okruženju e-trgovine. Za razliku od sustava temeljenih na pretraživanju nestrukturiranih dokumenata, predloženi agentski sustav koristi LLM-ove za interpretaciju poslovnih upita pisanih prirodnim jezikom, generiranje SQL-a i sintezu korisnih poslovnih uvida izravno iz relacijskih baza podataka. Sustav je dizajniran kao suradnja više specijaliziranih agenata (swarm), s agentima za generiranje SQL-a i analitičku interpretaciju, pri čemu se koriste najnapredniji otvoreni LLM-ovi poput Mistral 7B, Devstral, Qwen3 i Qwen2.5-Coder. Evaluacija se provodi pomoću automatiziranih semantičkih metrika (FLEX) i stručne ljudske procjene, uspoređujući jednostruke i višestruke agentske sustave prema točnosti, jasnoći i primjenjivosti uvida. Rezultati pokazuju da višestruki agentski sustavi, osobito oni koji kombiniraju Qwen2.5-Coder za SQL zadatke i Qwen3 za analizu, značajno nadmašuju jednostruke agente u tehničkoj ispravnosti i poslovnoj vrijednosti. Također ističe se važnost specijalizacije modela pomoću prompt engineering-a i suradničkih agentskih arhitektura za napredak automatizacije u domeni strukturiranih podataka.The exponential growth of data in e-commerce and digital marketing has created a pressing need for accessible, automated business analytics solutions. Traditional approaches to extracting insights from structured business data require significant technical expertise, particularly in formulating and executing SQL queries, which limits access to data-driven decision-making. This thesis investigates the deployment of open-source large language models (LLMs) as autonomous agents for SQL-based analytics in e-commerce environments. Unlike retrieval-augmented generation systems that rely on unstructured document retrieval, the proposed agentic framework leverages LLMs to interpret natural language business queries, generate executable SQL, and synthesize actionable business insights directly from relational databases. The system is architected as a multi-agent (swarm) collaboration, with specialized agents for SQL generation and analytical interpretation, each powered by state-of-the-art open-source LLMs including Mistral 7B, Devstral, Qwen3, and Qwen2.5-Coder. Evaluation is conducted using both automated semantic metrics (FLEX) and expert human judgment, comparing single-agent and multi-agent designs across accuracy, clarity, and actionability of insights. Results demonstrate that multi-agent systems, particularly those combining Qwen2.5-Coder for SQL tasks and Qwen3 for analysis, significantly outperform single-agent baselines in both technical correctness and business value. The findings highlight the importance of model specialization, prompt engineering, and collaborative agent architectures for advancing automated business intelligence in structured data domains
Optimization of gravitational force simulation with a large number of particles
No full textUsporedba brzine izvođenja dva načina izračuna gravitacijskih sila među tijelima u grafičkoj simulaciji.
U ovom radu uspoređuje se naivni algoritma izračuna gravitacijskog utjecaja koji ima kvadratnu vremensku složenost s Barnes-Hut algoritmom koji ima loglinearnu vremensku složenost. Isto tako se objašnjava način paralelizacije oba algoritma i na procesoru, ali i na grafičkoj kartici. Krajnje analiziraju se rezultati izvođenja svakog od algoritma pri različitim konfiguracijama kao što su izvođenja procesoru ili na grafičkoj kartici i dodatno broj za različit broj dretvi na procesoru. Opisano je i cjelokupno programsko rješenje koje uključuje i prikaz i interaktivno sučelje.A comparison of the execution speed of two methods for calculating gravitational forces between bodies in a graphical simulation.
In this paper, the naive algorithm for calculating gravitational influence, which has quadratic time complexity, is compared with the Barnes-Hut algorithm, which has logarithmic-linear time complexity. Also, the method of parallelization of both algorithms is explained, both on the processor and on the graphics card. Finally, the results of executing each of the algorithms under different configurations are analyzed, such as execution on the processor or on the graphics card, and additionally for different numbers of threads on the processor. The entire software solution is also described, which includes visualization and an interactive interface
Implementation of a decentralized savings and financing system using blockchain technology
No full textUobičajeno, ljudi se odlučuju na štednju u kreditnim institucijama poput oročene, dječje ili stambene štednje. Također, postoje opcije ulaganja u mirovinske ili investicijske fondove. Ipak, svaki od navedenih oblika ovisi o centraliziranom entitetu koji upravlja novcem štediša.
Korištenjem sustava implementiranog u radu, štednja ili obročno financiranje se može ostvariti eliminirajući potrebu za povjerenjem u centralizirane entitete, tako da sve funkcionira na predefinirani decentralizirani način.
Predstavljeno rješenje, temeljeno na sustavu ulančanih blokova, pruža alternativni pristup implementaciji određenih aspekata postojećeg financijskog sustava, nudeći veću fleksibilnost, uz niz prednosti, ali i određene izazove.Typically, people choose to save money in credit institutions through fixed-term savings, children's savings, or housing savings. There are also possibilities to invest in pension or investment funds. However, each of these forms relies on a centralized entity that manages the savers' funds.
By using the system implemented in this work, saving money or portion-based financing can be achieved without the need to trust centralized entities, as everything operates in a predefined decentralized way.
The presented solution, based on a blockchain system, offers an alternative approach to implementing certain aspects of the existing financial system, providing greater flexibility, along with a range of advantages, but also certain challenges
Ego-velocity estimation based on radar data
No full textTema rada je estimacija brzine na temelju podataka radara. Budući da radar neminovno uvijek ima lažne i pogešne detekcije, cilj je izbaciti što više takvih točaka prije same metode estimacije. U sklopu ovog rada to se pokušava postići time što se iz oblaka točaka radara filtriraju sve točke koje nisu na tlu, pod pretpostavkom da su točke na tlu češće statične i da su podatci izmjereni za te točke precizniji. Na kraju se ta metoda testira na javnom skupu podataka i uspoređuje s metodom koja se ne ograničava na točke tla.The topic of the thesis is ego-velocity estimation based on radar data. Since radar always has false and erroneous detections, the goal is to eliminate as many such
points as possible before applying the estimation method itself. This thesis attempts to
achieve this by filtering out all points that are not on the ground from the radar pointcloud, using the assumption that points on the ground are more often static and that the measurements for these points are more precise. Finally, this method is tested on a public dataset
and compared with a method that is not limited to ground points
Design and development of a low-cost floater for sustainable fishing
No full textWhile it is widely recognized that fish are an ecologically and commercially important group, our current knowledge of fish occurrence, composition (diversity), abundance and behavior (e.g. migration) is limited to anecdotal sightings and reports, often from laypersons. In situ marine monitoring bridges this gap and allows us to track and monitor marine life. One such system is the SOUND system: a swarm of low-cost Lagrangian floats that can non-invasively support aquaculture and fisheries, especially in remote areas and developing countries. The swarm of floats works together in a group and uses underwater acoustic communication. It provides long-term data on the fish population, which can shed light on the interdependencies of spatially segmented ecosystems, the top-down regulation of bio-geophysical processes and the sensitivity of the environment to anthropogenic stress factors. SOUND Floater consists of a piston-based buoy control system, an active sonar system with on-board analysis and a satellite communication module. It is capable of probing the water to a depth of 50 m while maintaining position with an accuracy of <10 cm, detecting schools of fish from a distance of 500 m and operating for 5 consecutive days. In this technical communication paper we present the detailed design of the SOUND prototype, including its mechanical, electrical and algorithmic parts. We report on results from laboratory pool and from two sea trials