ABSTRACT. Numerical algebraic geometry provides a toolbox of numerical methods for performing computations in algebraic geometry. Even though many computations which are performed on a computer using floating-point arithmetic are not certified, they can often be made very reliably using adaptive precision computations. Moreover, there is a wealth of information regarding the original problem which can be extracted from various numerical computation that can be used to improve subsequent symbolic computations to certify the result. This talk will highlight some recent successes of such hybrid numeric-symbolic methods in algebraic geometry.

ABSTRACT. Predicting cloud movement and dynamics is an important aspect in several areas, including prediction of solar energy generation. Knowing where a cloud will be or how it evolves over a given geographical area can help energy providers to better estimate their production levels. In this paper we propose a novel approach to predicting cloud movement based on satellite imagery. It combines techniques of generating motion vectors from sequential images with neural networks. First, the images are masked to isolate cloud pixels, then Farneback’s version of the Optical Flow algorithm is used to detect motion from one image to the next and generate motion vector flow for each pair of images. After that, a feed forward back propagation neural network is trained with the vector data derived from the dataset imagery. Different parameters for the duration of the training, size of the input, and the neighborhood radius of one point in the scene are used. Promising results are presented and discussed to weight the potential of the proposed algorithm for forecasting cloud cover and cloud position in a scene.

ABSTRACT. This paper presents two approaches of model order reduction applied to two different type of electrical motors, the Permanent Magnet Synchronous Reluctance Machine (PMASynRM) and the Switched Reluctance Motor(SRM). In the field of Electrical Machines, a motor can be described using a mathematical model, with complex non-linear differential equations, based on equivalent electric circuit parameters (inductances and resistances). Different theoretical and experimental methods have been proposed for estimating the inductances, requiring time consuming tests or simulations. Finding methods to reduce the number of simulations/measurements necessary to compute the parameters of the motors represents a constant concern in the Industry research field. Less measurements/simulations means reducing the computation time, which is a priority in Industry, for a shorter timeto-market. In our experiments we have chosen to reduce the problem dimensions for the computation of the magnetization characteristic of the machines, using Machine Learning. We compared Principal Component Analysis with Polynomial Interpolation and we have reduced the problem space with 50% up to 80%, depending on the motor type and context.

ABSTRACT. Cyber-Security industry has always been a "cat and a mouse" game - whenever a new technology was developed it was shortly followed by the appearance of several techniques used by malware creators to avoid detection. It is no surprise that the developing of adversarial machine learning algorithms has provided a tool that can be used to avoid machine learning based detection mechanisms available in security products.

This paper presents how the same algorithms can also be used to strengthen a security solution by identifying its weak points / features. We will also provide a method that can be used to fight GANs (Generative Adversarial Networks) with GANs that is effective when a malware writer is using these methods to avoid detection.

ABSTRACT. With the increase in malware samples in the last decade more antivirus products started to use machine learning algorithms in order to cope with the large volume of data. Thanks to the good results and advances in learning infrastructure the neural networks have become one of the preferred way of addressing this. However, these algorithms need to be fine tuned in order to not add an overhead of costly false positives.

This paper presents a study that takes a closer look into two techniques used for false positive mitigation issue: one side training and weight class adjustment. The techniques are used to train a neural network with zero false positives and are compared in order to find out which one give the highest true positive rate.

Using a large dataset constructed over several years we show that by using these techniques a 90% true positive rate can be obtained while training for 0 false positives.

ABSTRACT. We discuss building square, integer matrices, such that the Gram-Schmidt process leads to rational orthonormal matrices. We show some interesting properties that arise while searching for integer entries, as well as a process for doing this by direct computation.

ABSTRACT. Differential Evolution (DE) is arguably one of the most powerful stochastic optimization algorithms of current interest. Since its inception in 1995, DE has drawn the attention of many researchers all over the world resulting in a lot of variants of the basic algorithm with improved performance, especially on continuous parametric spaces. This talk will begin with a brief but comprehensive overview of the basic concepts related to DE, its algorithmic components and control parameters. It will subsequently discuss some of the significant algorithmic variants of DE for bound-constrained single-objective optimization for high-dimensional search spaces. The talk will then focus on some interesting DE variants with additional mechanisms like a distance-based selection, a clustering procedure and bi-objective formulations for solving multi-peak optimization problems where the objective is to locate all the global and local optima of a fitness landscape during one run of the algorithm. The talk will finally highlight a few open research problems in the related areas.

ABSTRACT. In this paper we introduce Competitive Unsupervised GrowCut, a cellular automata-based, unsupervised and autonomous algorithm that combines the label merging component of Unsupervised GrowCut with the soft label propagation mechanism of GrowCut. We evaluate our algorithm on two benchmark image segmentation datasets, along with two related methods proposed in the literature.We also provide a detailed comparative analysis of the three algorithms’ segmentation performance and properties. Our analysis identifies application-specific regimes that govern the relative performance of the analyzed algorithms.

ABSTRACT. Cellular automata have been widely used for solving the edge detection problem. This paper proposes an algorithm which optimizes cellular automata rules using Particle Swarm Optimization based on an existing method in the literature. Moreover, the method is extended from grayscale to RGB images by performing the optimization on each colour channel individually. A discussion on choosing the proper fitness function as well as comparative results with respect to the state-of-the-art are presented. As our algorithm is comparable to the Canny and Sobel edge detectors, it could be used in image segmentation tasks as a subroutine for edge detection.

ABSTRACT. One important image processing task is the edge detection in intensity images. Some approaches were proposed based on computing the gradient of the signal, but not many approaches were proposed based on Cellular Automata (CA). In this paper, an edge detection method based on CA evolved by the means of an Evolutionary Algorithm (EA) for grey images is presented. Promising results are obtained when compared with similar approaches in the specialized literature.

ABSTRACT. Extreme learning machine (ELM) is a non-iterative algorithm for training single-hidden layer feedforward neural network (SLFN). ELM has been shown to have good generalization performance and faster learning speed than conventional gradient-based learning algorithms. However, due to the random determination of the hidden neuron parameters (i.e., input weights and biases) ELM may require large number of neurons in the hidden layer. In this paper, the original harmony search (HS) and its variants, namely improved harmony search (IHS), global-best harmony search (GHS) and intelligent tuned harmony search (ITHS) are used to optimize the input weights and hidden biases of ELM. The output weights are analytically determined using the Moore-Penrose (MP) generalized inverse. The performance of the hybrid approaches is tested on several benchmark classification problems. The simulation results show that the integration of HS algorithms with ELM has obtained compact network architectures with good generalization performance

ABSTRACT. This paper explores the possibility of implementing an investment system that implements urban renewable energy by matching the investor resources, the technical private sector and real-estate owners. In this paper we postulate that there is a mechanism that provides benefits to all involved while automatically keeping profit margins fair. We analyze the options of renewable power generation installations in urban areas concluding that solar-wind arrays are most suited. We describe a mathematical model and algorithmic mechanism that matches owners with technical installers and investors. We proceed to analyze the model mathematically. Conclusions are drawn and system fairness is discussed. This system is presented as an alternative to bank loans by sharing the risk between three well intended agents while overseen by an automated system of rules and matching strategies. At the end of the paper we present a practical study case for Craiova, Romania, a temperate climate city from Eastern Europe. The final conclusions of the paper is that the system reaches an automated fair distribution of profit to all agents while presenting financial interest to the owner of the property.

ABSTRACT. This paper introduces a multi-agent system to enhance virtual network functions placement. The suggested system consists of a number of proactive agents aiming at providing a virtual network functions configuration depending on the state of the environment. Some of these agents are developed to determine an adequate target configuration based on available information and after their negotiation interaction; while other agents are designed to perceive the environment, as well as to act as managers in the decision process. The communication and the cooperation of these agents allow efficient placement of virtual network functions on distributed data-centers, possibly using multiple agent platforms. The results have demonstrated that our approach can enhance the solution by providing an adequate and automatic configuration without the intervention of any human decision maker.