Rade Božović

Computer Science and Information Systems (ComSIS)

Artificial Neural Network Modeling for Air Pollution Prediction: LSTM versus Levenberg-Marquardt Approach

Goran Keković; Rade Božović; Sonja Ketin; Vladimir Mikić; Miloš Ilić; Boban Vesin

In Press

Accurate prediction of air pollutant concentrations remains a critical challenge for environmental monitoring and public health, demanding robust and adaptive artificial intelligence approaches. This study investigates the effectiveness of various types of artificial neural networks (ANNs), including Long-Short Term Memory Networks (LSTM) and networks based on the Levenberg-Marquardt algorithm (LM) and its variant with Bayesian regularization (LMBR), in forecasting air pollution under different data conditions. The artificial neural networks were tested on two datasets: the Air quality dataset, where the target variable was the concentration of benzene (C6H6), and the Beijing PM2.5 dataset, where the target was the concentration of PM2.5 particles. The results indicate that LSTM networks are more robust in cases with non-monotonic relationships between predictors and target values, while the input variable selection technique (IVS) can be used to detect seasonal trends in the input data.

Artificial Intelligence; LSTM network; Levenberg-Marquardt algorithm; Input variable selection

M22

Udžbenik

Digitalna obrada signala

Rade Božović; Vladimir Orlić

—

—

—

—

3rd International Scientific and Professional Conference ALFATECH - Smart Cities and Modern Technologies 2026

Reframing Air Pollution Prediction: The Regression-Classification Dichotomy in Artificial Neural Networks

Goran Keković; Vladimir Mikić; Miloš Ilić; Rade Božović

https://conference.alfatech.rs/proceedings/

The problem of air pollution not only affects large urban areas and human health, but also threatens the environment in terms of danger to flora and fauna, water, soil and ecosystems, and affects climate change. Within real-time air pollution prediction and forecasting methods, artificial neural networks (ANN) are at the forefront. This paper examines the possibility of using Long-Short Memory (LSTM) neural networks as classifiers in real-time air pollution prediction tasks with PM2.5 particles. The results indicate that the regression method is effective in cases of air pollution assessment, while the classification method is suitable as an alarm in cases of air pollution dangerous to human life.

Air pollution; LSTM networks; PM2.5 particles; Regression method; Classification method

M31

3rd International Scientific and Professional Conference ALFATECH - Smart Cities and Modern Technologies 2026

Communication Scalability Analysis of Federated Learning in Autonomous Vehicle Networks

Vladimir Mladenović; Vladimir Orlić; Rade Božović

https://conference.alfatech.rs/wp-content/uploads/2026/05/Alfatech-book-of-abstracts-2026-1.pdf

Autonomous vehicles generate and exchange large volumes of data used to improve artificial intelligence models. Traditional centralized training approaches face challenges related to scalability, data privacy, and network communication load. Federated learning offers a potential solution by enabling distributed model training without transferring raw data. This paper analyzes the communication scalability of federated learning in autonomous vehicle networks. A simulation model is developed in which autonomous vehicles communicate with a central aggregation server using bidirectional TCP communication. The results provide insight into communication limitations in such systems and represent a basis for further research focused on optimizing communication protocols and improving distributed learning architectures.

Autonomous Vehicles; Federated Learning; Communication Scalability; Distributed Machine Learning

M31

Radioengineering

Artificial bias induction in fourth-order cumulants based automatic modulation classification algorithm in AWGN and multipath propagation channel

Rade Božović; Vladimir Orlić; Goran Keković

https://www.radioeng.cz/fulltexts/2025/25_02_0224_0233.pdf

Automatic modulation classification (AMC) predstavlja široko primenjenu tehniku za prepoznavanje formata modulacije signala koji se unapred ne poznaju. Zbog male algoritamske i hardverske složenosti, AMC algoritmi zasnovani na kumulantima četvrtog reda i dalje su veoma popularni. Prisustvo pristrasnosti u standardnim procenjenim vrednostima kumulanata realnih signalnih konstelacija pozitivno utiče na rezultat klasifikacije pri razlikovanju realnih i kompleksnih signala. Zbog toga je u radu predložen novi pristup AMC-u, sa fokusom na manipulaciju teorijskim očekivanim vrednostima kumulanata realnih signalnih konstelacija, uz pretpostavku da veštački uvedena pristrasnost može poboljšati performanse AMC-a. Veštačko uvođenje pristrasnosti realizovano je modifikacijama standardne matematičke formule kumulanata. Performanse modifikovanih i standardnih AMC algoritama zasnovanih na kumulantima četvrtog reda ispitane su za realne i kompleksne signalne konstelacije kroz Monte Karlo simulacije u uslovima propagacije sa aditivnim belim Gausovim šumom (AWGN) i višestaznim kanalom, sa poznatim i nepoznatim impulsnim odzivom. Evaluacija je sprovedena preko verovatnoće tačne klasifikacije. Prikazani numerički rezultati potvrdili su superiornost algoritma zasnovanog na veštački indukovanoj pristrasnosti u klasifikaciji realnih i kompleksnih signala, u svim razmatranim scenarijima propagacije, naročito u radio-okruženju sa nižim vrednostima odnosa signal-šum (SNR). Značajna poboljšanja performansi AMC-a dostižu i do 25%.

AMC; AWGN; bias; Binary Phase Shift Keying (BPSK); channel impulse response; cumulants; multipath; Quadrature Amplitude Modulation (QAM)

M23

International Scientific and Professional Conference – ALFATECH Smart Cities and Modern Technologies

Optimization of AI Methods for Air Pollution Prediction

Goran Keković; Rade Božović

Optimization of AI Methods for Air Pollution Prediction – DOI UBKG

Jedan od najvećih problema velikih urbanih sredina jeste zagađenje vazduha, a metode veštačke inteligencije (AI) mogu predviđati nivo zagađenja koristeći širok spektar parametara. U ovom radu razmatra se primena veštačkih neuronskih mreža (ANN) zasnovanih na Levenberg–Marquardt algoritmu sa Bajesovom regularizacijom (LMBR). Pokazano je da ovaj algoritam ostvaruje veoma visoku tačnost predikcije, konkurentnu radial basis neuronskim mrežama, koje se često koriste za regresione zadatke. Takođe je pokazano da se izborom optimalne veličine uzorka, uz podešavanje parametara ANN-a, može postići ravnoteža između željene tačnosti metode i odstupanja između simuliranih i realnih podataka. Relativna greška korišćena je kao mera tog odstupanja. Istovremeno je pokazano da veličina uzorka nije uvek presudan faktor koji utiče na efikasnost AI metode, već da se potpuna slika dobija kada se uzme u obzir celokupna struktura ulaznih podataka.

Air pollution; Artificial neural networks; Bayesian regularization; Levenberg–Marquardt algorithm

M33

11th International Scientific Conference on Defense Technologies (OTEH 2024)

Standard cumulant-based automatic modulation classification performance under colored noise channel conditions

Rade Božović; Vladimir Orlić

https://scindeks-clanci.ceon.rs/data/pdf/proc-0018/2024/proc-00182400415B.pdf

Automatic modulation classification (AMC) predstavlja široko primenjenu tehniku obrade signala u situacijama kada format modulacije primljenog signala nije unapred poznat. Od presudnog je značaja za različite vojne i komercijalne komunikacije. Zbog male složenosti algoritma, hardverskih zahteva i drugih praktičnih aspekata, algoritmi zasnovani na strukturama kumulanata četvrtog i šestog reda, odnosno statistici višeg reda, i dalje su konkurentna i aktuelna istraživačka tema. Degradacija signala usled prisustva šuma značajno narušava performanse AMC-a. U ovom radu analizirane su performanse standardnih AMC algoritama zasnovanih na kumulantima u kontekstu uticaja obojenog šuma. Analiza je sprovedena Monte Karlo simulacijama u Matlab softveru, u propagacionom okruženju sa kanalom obojenog šuma.

Automatic modulation classification; colored noise; cumulants; feature based

M33

The First International Scientific Conference (Alfatech – Smart Cities and Technologies 2024)

An optimal threshold adaptation of energy detector in cognitive radio for smart cities applications

Rade Božović; Goran Keković

https://conference.alfatech.rs/wp-content/uploads/2025/05/AN-OPTIMAL-THRESHOLD-ADAPTATION-OF-ENERGY-DETECTOR-IN-COGNITIVE-RADIO-FOR-SMART-CITIES-APPLICATIONS.pdf

Pametan grad predstavlja savremeni tehnološki koncept međusobnog povezivanja različitih uređaja sa ciljem prikupljanja i naknadne obrade različitih tipova podataka. Na osnovu tako dobijenih informacija optimizuje se i unapređuje efikasnost različitih gradskih servisa, kao što su bezbednost, potrošnja energije i sigurnost, što značajno podiže kvalitet života građana. Ovaj koncept podrazumeva integraciju naprednih informaciono-komunikacionih tehnologija. Nedostatak radio-spektra jedan je od glavnih izazova za performanse bežičnih mreža zbog problema pokrivenosti i kapaciteta. Tehnologija kognitivnog radija omogućava efikasno korišćenje radio-spektra. Da bi se izbegla ili minimizovala interferencija, funkcija spektralnog osluškivanja ima ključnu ulogu u radu kognitivnog radija. Zbog niske složenosti implementacije, nekoherentni energetski detektor, zasnovan na poređenju detektovane energije signala sa pragom odlučivanja, predstavlja dobar izbor za spektralno osluškivanje. U ovom radu opisan je zatvoreni oblik rešenja za adaptaciju praga energetskog detektora u kognitivnom radiju.

Advanced smart communications; cognitive radio; energy detector; spectrum sensing; threshold adaptation

M33

The First International Scientific Conference (Alfatech – Smart Cities and Technologies 2024)

A modified version of gradient descent algorithm as a solution of local minimum problem in artificial neural network

Goran Keković; Rade Božović; Negovan Stamenković

https://conference.alfatech.rs/wp-content/uploads/2025/05/AN-OPTIMAL-THRESHOLD-ADAPTATION-OF-ENERGY-DETECTOR-IN-COGNITIVE-RADIO-FOR-SMART-CITIES-APPLICATIONS.pdf

U ovom radu predložen je softver zasnovan na modifikovanom algoritmu gradijentnog spusta i algoritmu propagacije unazad (backpropagation) za prevazilaženje problema lokalnog minimuma u veštačkim neuronskim mrežama. Tokom obuke veštačke neuronske mreže, na kraju svake epohe proveravano je postojanje globalnog minimuma kroz sukcesivne vrednosti funkcije gubitka, kao i određivanjem procenta uspešno klasifikovanih uzoraka iz trening i test skupova. Softver je napisan u programskom jeziku C# na objektno orijentisan način. Realizovan je modularno, u smislu da poseduje sopstvenu matematičku biblioteku i može se nadograđivati drugim algoritmima veštačke neuronske mreže.

Artificial neural networks; Neuron weights; Global minimum; Gradient descent; Backpropagation algorithm; C# programming language

M33

10th International Scientific Conference on Defense Technologies (OTEH 2022)

Signal constellation distortion and its impact on cumulant-based AMC performance

Vladimir Orlić; Rade Božović

http://www.vti.mod.gov.rs/oteh22/elementi/rad/053.pdf

Automatic modulation classification (AMC) od ključnog je značaja za različite vojne i komercijalne komunikacije, u kojima signalna konstelacija nije unapred poznata na prijemu. Zbog svoje jednostavnosti, među brojnim do sada razvijenim AMC algoritmima, algoritmi zasnovani na strukturama kumulanata višeg reda ostaju konkurentni u pogledu praktične primenljivosti i zbog toga su u fokusu istraživača širom sveta. Ipak, većina istraživanja koja se posebno bave ovim algoritmima uglavnom je fokusirana na aditivni beli Gausov šum, višestazno prostiranje i interferenciju kao izvore degradacije signala, kao i na odgovarajuće AMC performanse. U ovom radu po prvi put analiziran je skup praktičnih izvora distorzije signalne konstelacije u kontekstu AMC-a, i to: neuravnoteženost amplitude, neuravnoteženost faze i prisustvo faznog džitera. Uticaj ovih efekata prikazan je na performanse AMC-a u slučajevima standardnih kumulanata četvrtog reda, standardnih kumulanata šestog reda i nepristrasnih kumulanata šestog reda, u uslovima šuma, uz potvrdu razmatranja kroz veći broj Monte Karlo simulacija.

Automatic modulation classification; cumulants; phase jitter; amplitude imbalance; phase imbalance

M33

2021 International Conference on Computational Performance Evaluation (ComPE)

Estimation of bias in numerical values of normalized sixth-order cumulants’ structures for various signal constellations

Rade Božović; Vladimir Orlić

https://ieeexplore.ieee.org/document/9752436

Automatic modulation classification (AMC) is a technique for modulation format recognition of signals considered to be a priori unknown and has wide usage in numerous wireless systems and applications. Due to low algorithm complexity and other practical aspects, AMC algorithms based on higher-order statistics are very popular. In this paper, the properties of AMC algorithms based on sixth-order cumulants are explored in the context of real and complex signals, i.e. different Pulse Amplitude Modulation (PAM) and Quadrature Amplitude Modulation (QAM) signal constellations. The analysis was performed through Monte Carlo simulations in propagation conditions with Additive White Gaussian Noise (AWGN) and multipath propagation, with a focus on numerical quantification of the presence of bias in standard cumulant structures. A recently published AMC approach with unbiased classification features was tested in the same context, confirming that bias issues can be resolved with the new sixth-order cumulants formula and that better classification performance of real and complex signals can be ensured in the considered propagation scenarios.

Automatic modulation classification; AWGN; cumulants; multipath; PAM; QAM; bias

M33

2021 29th Telecommunication Forum (TELFOR)

Impact of AWGN estimation on classification performance of AMC algorithms based on higher order cumulants

Marko Nerandžić; Rade Božović; Vladimir Orlić

https://ieeexplore.ieee.org/document/9653393

Automatic modulation classification (AMC) is a technique for modulation format recognition of a priori unknown signals, with wide usage in numerous wireless systems and applications. In this paper, the performance of AMC algorithms based on higher-order cumulants was explored, since low computational complexity makes them attractive from a practical point of view. To further explore practical aspects of these algorithms, they were considered under conditions of Additive White Gaussian Noise (AWGN) with unknown noise variance, in a channel with AWGN only, and in a multipath channel. Monte Carlo simulations were executed for performance tests, with AMC features expressed through probability of correct classification of modulation formats. The impact of using estimated noise variance, instead of a priori known noise variance, on the performance of AMC algorithms was presented and discussed, confirming their potential for practical application.

AMC; AWGN; cumulants; estimation; variance

M33