Browsing Computer Science by Title
Prethodna
Strana 36-40 od 40
-
Bačanin Džakula, Nebojša (Beograd , 2015)[more][less]
Abstract: Hard optimization problems that cannot be solved within acceptable computational time by deterministic mathematical methods have been successfully solved in recent years by population-based stochastic metaheuristics, among which swarm intelligence algorithms represent a prominent class. This thesis investigates improvements of the swarm intelligence metaheuristics by hybridization. During analysis of the existing swarm intelligence metaheuristics in some cases de ciencies and weaknesses in the solution space search mechanisms were observed, primarily as a consequence of the mathematical model that simulates natural process as well as inappropriate balance between intensi cation and diversi cation. The thesis examines whether existing swarm intelligence algorithms for global optimization could be improved (in the sense of obtaining better results, faster convergence, better robustness) by hybridization with other algorithms. A number of hybridized swarm intelligence metaheuristics were developed and implemented. Considering the fact that good hybrids are not created as a random combination of individual functional elements and procedures from di erent algorithms, but rather established on comprehensive analysis of the functional principles of the algorithms that are used in the process of hybridization, development of the hybrid approaches was preceded by thorough research of advantages and disadvantages of each involved algorithm in order to determine the best combination that neutralizes disadvantages of one approach by incorporating the strengths of the other. Developed hybrid approaches were veri ed by testing on standard benchmark sets for global optimization, with and without constraints, as well as on well-known practical problems. Comparative analysis with the state-of-the-art algorithms from the literature demonstrated quality of the developed hybrids and con rmed the hypothesis that swarm intelligence algorithms can be successfully improved by hybridization. URI: http://hdl.handle.net/123456789/4245 Files in this item: 1
phdBacaninNebojsa.pdf ( 3.813Mb ) -
Alatrash, Emhimed Salem (Beograd , 2015)[more][less]
Abstract: Ontologies, often defined as an explicit specification of conceptualization, are necessary for knowledge representation and knowledge exchange. This means that ontology describes concepts and relations that exist in a domain. To enable knowledge exchange, it is necessary to describe these concepts and relations in a better way than just ordering them in taxonomy. A computational ontology consists of a number of different components, such as Concepts, Instances, Individuals or Facts, Relations and Attributes. The present research is intended to consider different software tools related to Semantic web, and achieve a kind of comparison among them. In fact, five ontology-editors are described and compared. They are: Apollo, Onto Studio, Protégé, Swoop and TopBraid Composer Free Edition. The structure and basic features of these editors as well as the way of using them are described. The main criterion used in the process of comparing these editors lies in their convenience for the user, and the possibility to apply them in different kinds of application. The main goal of the work is to introduce a method for ontology construction of a certain domain in applying the Semantic web. A number of software tools adapted to build up the domain ontologies of most wide–spread natural languages are available; however accomplishing that for any given natural language presents a challenge. This research proposes a semi-automatic procedure to create ontologies for different natural languages. The approach utilizes various software tools that are available on the Internet, most notably DODDLE-OWL which is a domain ontology development tool implemented for English and Japanese languages. Through this tool, WordNet, Protégé and XSLT transformations, the researcher proposes a general procedure to construct domain ontology for any natural language. URI: http://hdl.handle.net/123456789/4266 Files in this item: 1
phdEmhimedAlatrash.pdf ( 2.171Mb ) -
Nikolić, Mladen (Belgrade , 2013)[more][less]
Abstract: In this thesis the problem of guiding search in automated theorem proving is considered. The thesis consists of two parts that have the CDCL search system, the system intensively used by modern SAT solvers, as their common topic. In the rst part of the thesis a simple approach to guiding search is considered | guiding by the selection of the solver, its heuristics, and their parameters, based on the properties of an instance to be solved. The basis of the proposed methods for algorithm selection is syntactical similarity of formulae which is re ected in their graph structure. This graph similarity is established and analyzed by using an original graph similarity measure (which turned out to be useful in other contexts, too). Yet, practical approaches to measuring similarity of formulae are based on their numerical features due to the computational complexity issues. Two simple methods for algorithm selection, based on k nearest neighbors, were proposed. The rst technique, ArgoSmArT is based on classi cation of instance in one of the prede ned families for which the e cient algorithms are known. The instance is solved by algorithm corresponding to the family to which the instance was classi ed. The second technique, ArgoSmArT k-NN is based on nding several similar instances in the training set for which the solving times by all considered algorithms are known. The instance is solved by the algorithm that behaves the best on those instances. ArgoSmArT technique is better suited for con guration selection of a SAT solver, and ArgoSmArT k-NN for SAT solver selection. ArgoSmArT k-NN technique showed to be more e cient than the most important and very complex system for SAT solver selection | SATzilla system. Apart from CNF SAT solver selection, the problem of non-CNF SAT solver selection is considered. The focus was not on solver selection techniques, since the proposed techniques are directly applicable, but on the attributes that can be used to describe non-CNF SAT instances, which have not been proposed earlier. The results in this domain are positive, but still limited. The main reason for that is the lack of greater number of non-CNF SAT solver of di erent behaviour, which is not surprising, having in mind that this kind of solvers is in its early stage of development. Apart from construction of e cient SAT solver selection system, the methodology of SAT solver comparison, based on statistical hypothesis testing is proposed. The need for such a methodology comes from great run time variations of single instance solving by a solver, which can result in di erent SAT solver orderings when one tries to compare their performance or rank them, as experimentally demonstrated. The proposed methodology gives the estimate of statistical signi cance of the performed test and the estimate of the e ect size, for instance the probability of a solver being faster than another. The second part of the thesis is concerned with generalizing CDCL search system to fragments of rst order logic. The proposed system can be used as a basis for e cient proving in some fragment if the rules of resolution and factoring are speci ed for that fragment. These rules are de ned for an extension of coherent logic. The soundness and completeness of the system are proved. The system has several distinguishing features which are a consequence of previously performed analysis of challenges in coherent logic theorem proving. The system enables rst order reasoning, instead of ground one characteristic for all existing coherent logic provers. Moreover, it introduces backjumps and lemma learning. The special attention in system design was paid to the possibility of generating readable proofs by the prover implementing the system. This possibility is one of the greatest qualities of coherent logic, but it is not easy to achieve if CDCL search system is used. One of the properties of the system that came from the need for generation of readable proofs is preservation of quanti ers in proving process which is rather unusual for existing CDCL systems. Another advantage of the proposed CDCL system is the possibility of transfer of heuristics which are already successfully employed in SAT solving to other domains. Based on the proposed system, the proof procedure Calypso for extended coherent logic was de ned which can also be used in standard coherent logic. The extension of Rete algorithm which enables detection of con icts and literals to be propagated or decided is proposed. Procedure Claypso is implemented in C++. It was evaluated on a representative coherent logic problems and it showed superior to other coherent logic provers and also the prover Vampire, the most e cient prover for rst order logic. Based on the results presented in this thesis, it can be concluded that the main hypothesis of this work, that the search system used in CDCL SAT solvers can be signi cantly improved by simple guiding and that it can be successfully formulated for fragments of rst order logic such as coherent logic, was con rmed and that the concrete answers on how to do that were provided. URI: http://hdl.handle.net/123456789/2584 Files in this item: 1
nikolic_mladen.pdf ( 1.448Mb ) -
Dimovski, Igor (Novi Sad , 2011)[more][less]
Abstract: A comprehensive pedagogical research regarding teaching mathematics at a tertiary, university level has been presented in the PhD dissertation. The educational resources tailored in an electronic form using the programme package Matlab are integrated in the learning process. The impact of ICT use to the essential knowledge that refers to multivariate calculus (functions of several variables, vector-valued functions and the three-dimensional analytical geometry) has been statistically explored by intensive use of 3D static and dynamic visual tools. Part of the students who have participated in the research have developed Matlab programmes all by their own. One part of the research has been focused on probable impact of the programming skills on learning mathematical concepts. URI: http://hdl.handle.net/123456789/3874 Files in this item: 1
PhD_I_Dimovski.pdf ( 5.423Mb ) -
Protić, Danijela (Beograd , 2023)[more][less]
Abstract: Anomaly detection is the recognition of suspicious computer network behavior by comparing unknown network traffic to a statistical model of normal network behavior. Binary classifiers based on supervised machine learning are good candidates for normality detection. This thesis presents five standard binary classifiers: the k-nearest neighbors, weighted k-nearest neighbors, decision trees, support vector machines and feedforward neural network. The main problem with supervised learning is that it takes a lot of data to train high-precision classifiers. To reduce the training time with minimal degradation of the accuracy of the models, a two-phase pre-processing step is performed. In the first phase, numeric attributes are selected to reduce the dataset. The second phase is a novel normalization method based on hyperbolic the tangent function and the damping strategy of the Levenberg-Marquardt algorithm. The Kyoto 2006+ dataset, the only publicly available data set of real-world network traffic intended solely for anomaly detection research in computer networks, was used to demonstrate the positive impact of such pre-processing on classifier training time and accuracy. Of all the selected classifiers, the feedforward neural network has the highest processing speed, while the weighted k-nearest neighbor model proved to be the most accurate. The assumption is that when the classifiers work concurrently, they should detect either an anomaly or normal network traffic, which occasionally is not the case, resulting in different decision about the anomaly, i.e. a conflict arises. The conflicting decision detector performs a logical exclusive OR (XOR) operation on the outputs of the classifiers. If both classifiers simultaneously detected an anomaly or recognized traffic as normal, their decision was no conflict had occurred. Otherwise a conflict is detected. The number of conflicts detected provides an opportunity for additional detection of changes in computer network behavior. URI: http://hdl.handle.net/123456789/5599 Files in this item: 1
Danijela Protic - Doktorska Disertacija.pdf ( 3.143Mb )
Prethodna
Strana 36-40 od 40