Milena Petkovic

Dr. Milena Petkovic


Research Focus & Metrics

Academic Career Length

17Years (2010-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
27 October 2020 in Journal of Global Optimization
Verified
Journal Article
Open access
PDF file
ARAnsgar Rössig
Ansgar Rössig
MPMilena Petkovic
Milena Petkovic
We consider the problem of verifying linear properties of neural networks. Despite their success in many classification and prediction tasks, neural networks may return unexpected results for certain inputs. This is highly problematic with respect to the application of neural networks for safety-critical tasks, e.g. in autonomous driving. We provide an overview of algorithmic approaches that aim to provide formal guarantees on the behaviour of neural networks. Moreover, we present new theoretical results with respect to the approximation of ReLU neural networks. On the other hand, we implement a solver for verification of ReLU neural networks which combines mixed integer programming with specialized branching and approximation techniques. To evaluate its performance, we conduct an extensive computational study. For that we use test instances based on the ACAS Xu system and the MNIST handwritten digit data set. The results indicate that our approach is very competitive with others, i.e. it outperforms the solvers of Bunel et al. (in: Bengio, Wallach, Larochelle, Grauman, Cesa-Bianchi, Garnett (eds) Advances in neural information processing systems (NIPS 2018), 2018) and Reluplex (Katz et al. in: Computer aided verification—29th international conference, CAV 2017, Heidelberg, Germany, July 24–28, 2017, Proceedings, 2017). In comparison to the solvers ReluVal (Wang et al. in: 27th USENIX security symposium (USENIX Security 18), USENIX Association, Baltimore, 2018a) and Neurify (Wang et al. in: 32nd Conference on neural information processing systems (NIPS), Montreal, 2018b), the number of necessary branchings is much smaller. Our solver is publicly available and able to solve the verification problem for instances which do not have independent bounds for each input neuron.

1 September 2012 in Expert Systems with Applications
Verified
Journal Article
Closed access
MPMilena Petkovic
Milena Petkovic
MRMilan Rade Rapaić
Milan Rade Rapaić
ZJZoran D. Jeličić
Zoran D. Jeličić
APAlessandro Pisano
Alessandro Pisano

28 June 2021 in Energy Science & Engineering
Verified
Journal Article
Open access
PDF file
MPMilena Petkovic
Milena Petkovic
TKThorsten Koch
Thorsten Koch
JZJanina Zittel
Janina Zittel
Germany is the largest market for natural gas in the European Union, with an annual consumption of approx. 95 billion cubic meters. The German high‐pressure gas pipeline network's length is roughly 40 000 km, which enables highly fluctuating quantities of gas to be transported safely over long distances. Considering that similar amounts of gas are also transshipped through Germany to other EU states, it is clear that Germany's gas transport system is essential to the European energy supply. Since the average velocity of gas in a pipeline is only 25 km/h, an adequate high‐precision, high‐frequency forecasting of supply and demand is crucial for efficient control and operation of such a transmission network. We propose a deep learning model based on spatio‐temporal convolutional neural networks (DLST) to tackle the problem of gas flow forecasting in a complex high‐pressure transmission network. Experiments show that our model effectively captures comprehensive spatio‐temporal correlations through modeling gas networks and consistently outperforms state‐of‐the‐art benchmarks on real‐world data sets by at least 15%. The results demonstrate that the proposed model can deal with complex nonlinear gas network flow forecasting with high accuracy and effectiveness.

21 November 2025 in Progress in Crystal Growth and Characterization of Materials
Verified
Journal Article
Closed access
MPMilena Petkovic
Milena Petkovic
LVL. Vieira
L. Vieira
Dropka NatashaDropka Natasha
Dropka Natasha

1 January 2016 in Srpski arhiv za celokupno lekarstvo
Verified
Journal Article
Open access
PDF file
AKAleksandar Knežević
Aleksandar Knežević
MPMilena Petkovic
Milena Petkovic
AMAleksandra Mikov
Aleksandra Mikov
MJMilica Jeremic-Knezevic
Milica Jeremic-Knezevic
CDCila Demesi-Drljan
Cila Demesi-Drljan
ZJZoran D. Jeličić
Zoran D. Jeličić
Introduction. Identification of predictive factors for walking ability with a prosthesis, after lower limb amputation, is very important in order to define patient?s potentials and realistic rehabilitation goals, however challenging they are. Objective. The objective of this study was to investigate whether variables determined at the beginning of rehabilitation process are able to predict walking ability at the end of the treatment using support vector machines (SVMs). Methods. This research was designed as a retrospective clinical case series. The outcome was defined as three-leveled ambulation ability. SVMs were used for predicting model forming. Results. The study included 263 patients, average age 60.82 ?} 9.27 years. In creating SVM models, eleven variables were included: age, gender, cause of amputation, amputation level, period from amputation to prosthetic rehabilitation, Functional Comorbidity Index (FCI), presence of diabetes, presence of a partner, restriction concerning hip or knee extension, residual limb hip extensor strength, and mobility at admission. Six SVM models were created with four, five, six, eight, 10, and 11 variables, respectively. Genetic algorithm was used as an optimization procedure in order to select the best variables for predicting the level of walking ability. The accuracy of these models ranged from 72.5% to 82.5%. Conclusion. By using SVM model with four variables (age, FCI, level of amputation, and mobility at admission) we are able to predict the level of ambulation with a prosthesis in lower limb amputees with high accuracy.

12 August 2021 in Energy Systems
Verified
Journal Article
Open access
PDF file
MPMilena Petkovic
Milena Petkovic
YCYing Chen
Ying Chen
IGInken Gamrath
Inken Gamrath
UGUwe Gotzes
Uwe Gotzes
NHNatalia Selini Hadjidimitrou
Natalia Selini Hadjidimitrou
TKThorsten Koch
Thorsten Koch
About 23% of the German energy demand is supplied by natural gas. Additionally, for about the same amount Germany serves as a transit country. Thereby, the German network represents a central hub in the European natural gas transport network. The transport infrastructure is operated by transmissions system operators (TSOs). The number one priority of the TSOs is to ensure the security of supply. However, the TSOs have only very limited knowledge about the intentions and planned actions of the shippers (traders). Open Grid Europe (OGE), one of Germany’s largest TSO, operates a high-pressure transport network of about 12,000 km length. With the introduction of peak-load gas power stations, it is of great importance to predict in- and out-flow of the network to ensure the necessary flexibility and security of supply for the German Energy Transition (“Energiewende”). In this paper, we introduce a novel hybrid forecast method applied to gas flows at the boundary nodes of a transport network. This method employs an optimized feature selection and minimization. We use a combination of a FAR, LSTM and mathematical programming to achieve robust high-quality forecasts on real-world data for different types of network nodes.