Fakulta informačních technologií
http://hdl.handle.net/10467/3566
FITThu, 23 May 2019 08:09:24 GMT2019-05-23T08:09:24ZVylepšení směrování v oportunistických sítích pomocí metod strojového učení s učitelem a bez učitele
http://hdl.handle.net/10467/81943
Vylepšení směrování v oportunistických sítích pomocí metod strojového učení s učitelem a bez učitele; Improvement of the Routing in Opportunistic Networks by the Application of Unsupervised and Supervised Machine Learning Techniques
Ladislava Smítková Janků
The dissertation thesis deals with the issue of special routing algorithms designed for com-munication in opportunistic networks. We proposed three extended routing algorithms combining unsupervised or supervised learning and existing routing methods. The oppor-tunistic communication networks are the ad-hoc networks where no assumption is made on the existence of a complete physical path between two nodes wishing to communicate. In opportunistic networks, the messages are transmitted when the node opportunistically meets another node; the characteristics of node movement can improve message transmis-sion. In practice, the network nodes can be mobile robots, wireless equipments carried by people, vehicles, wild animals, unmanned aerial vehicles. We have chosen three mutually independent problems to be solved. The rst problem is related to the applicability of unsu-pervised learning to the analysis of the data of node movements including implementation of the routing algorithm using the results of unsupervised learning. The second problem is related to the applicability of supervised learning and implementation of supervised learn-ing into a routing algorithm. The third problem is related to the enhancement of routing algorithms using unsupervised learning (statistical node mobility models) and the active node behavior. The active node behavior means that the node itself actively changes the route in order to move to the location more suitable for forwarding the message. In order to solve the rst problem, we proposed the Routing with Clustering algorithm. This method extends the standard epidemic routing with the limited message buer using knowledge extracted from the training data using clustering. In order to solve the second problem, we proposed a method combining the collection of labeled training data on message deliv-ery, a two class classier and an implementation of this classier into routing algorithm. In order to solve the third problem, we proposed the Active node movement algorithm (ANMA). Two versions of ANMA were designed. The rst version of the algorithm uses node densities in sectors. The second version of ANMA models data using GMRFs. Both versions of the algorithms make decisions about message forwarding and active node devi-ation from its planned route. All the proposed algorithms were tested on simulated data in the simulation environment. The experimental results were compared to the results obtained by the existing standard routing methods: Epidemic routing with the limited iii message buer and PRoPHET routing. All methods show a slight improvement in routing in opportunistic network from the viewpoint of a number of delivered messages or of the number of delivered messages in limited time. The signicant improvement was observed only in experiments using active node movement algorithm (ANMA), where nodes can improve routing using deviation from their planned routes.; The dissertation thesis deals with the issue of special routing algorithms designed for communication in opportunistic networks. We proposed three extended routing algorithms combining unsupervised or supervised learning and existing routing methods. The opportunistic communication networks are the ad-hoc networks where no assumption is made on the existence of a complete physical path between two nodes wishing to communicate. In opportunistic networks, the messages are transmitted when the node opportunistically meets another node; the characteristics of node movement can improve message transmission. In practice, the network nodes can be mobile robots, wireless equipments carried by people, vehicles, wild animals, unmanned aerial vehicles. We have chosen three mutually independent problems to be solved. The rst problem is related to the applicability of unsupervised learning to the analysis of the data of node movements including implementation of the routing algorithm using the results of unsupervised learning. The second problem is related to the applicability of supervised learning and implementation of supervised learning into a routing algorithm. The third problem is related to the enhancement of routing algorithms using unsupervised learning (statistical node mobility models) and the active node behavior. The active node behavior means that the node itself actively changes the route in order to move to the location more suitable for forwarding the message. In order to solve the rst problem, we proposed the Routing with Clustering algorithm. This method extends the standard epidemic routing with the limited message buer using knowledge extracted from the training data using clustering. In order to solve the second problem, we proposed a method combining the collection of labeled training data on message delivery, a two class classier and an implementation of this classier into routing algorithm. In order to solve the third problem, we proposed the Active node movement algorithm (ANMA). Two versions of ANMA were designed. The rst version of the algorithm uses node densities in sectors. The second version of ANMA models data using GMRFs. Both versions of the algorithms make decisions about message forwarding and active node deviation from its planned route. All the proposed algorithms were tested on simulated data in the simulation environment. The experimental results were compared to the results obtained by the existing standard routing methods: Epidemic routing with the limited iii message buer and PRoPHET routing. All methods show a slight improvement in routing in opportunistic network from the viewpoint of a number of delivered messages or of the number of delivered messages in limited time. The signicant improvement was observed only in experiments using active node movement algorithm (ANMA), where nodes can improve routing using deviation from their planned routes.
Tue, 30 Apr 2019 00:00:00 GMThttp://hdl.handle.net/10467/819432019-04-30T00:00:00ZAdaptivní algoritmy řízení inteligentních agentů
http://hdl.handle.net/10467/81942
Adaptivní algoritmy řízení inteligentních agentů; Adaptive control algorithms of intelligent agents
Martin Šlapák
The multi-agent systems and agent-based approach to complex and non-trivial practical problems have become more popular in recent years. The traditional methods of agent?s control sometimes fail or are inappropriate. The adaptive methods seem to be a possible solution for these cases. This thesis deals with an exploration of possibilities of design and development of adaptive methods of intelligent software agent?s control. These agents solve problems from data mining domain. The thesis brings an overview of given problem areas, describes experiments, and their results. In conclusions, the topics of author?s future research are proposed.; The multi-agent systems and agent-based approach to complex and non-trivial practical problems have become more popular in recent years. The traditional methods of agent?s control sometimes fail or are inappropriate. The adaptive methods seem to be a possible solution for these cases. This thesis deals with an exploration of possibilities of design and development of adaptive methods of intelligent software agent?s control. These agents solve problems from data mining domain. The thesis brings an overview of given problem areas, describes experiments, and their results. In conclusions, the topics of author?s future research are proposed.
Tue, 30 Apr 2019 00:00:00 GMThttp://hdl.handle.net/10467/819422019-04-30T00:00:00Z(Nelineární) stromové indexování a protisměrné vyhledávání
http://hdl.handle.net/10467/81824
(Nelineární) stromové indexování a protisměrné vyhledávání; (Nonlinear) Tree Pattern Indexing and Backward Matching
Jan Trávníček
Trees are one of the fundamental data structures used in Computer Science. The dissertation thesis contributions are best categorised as a part of arbology research [52]. Arbology research is a counterpart of stringology research. Arbology research deals with trees represented in some linear notations, i.e. like strings with additional properties that encode the tree structure. Many algorithms belonging to the stringology maybe, with some care, adapted to handle trees represented as strings using some linear notation. This dissertation thesis is focused on finding all occurrences of tree patterns and nonlinear tree patterns inside a subject tree. Two different general approaches of solving the problem are explored in the dissertation thesis. The first approach is focused on preprocessing of the subject tree and forming a complete index of the subject tree capable of reporting the occurrences when queried with (nonlinear) tree patterns. The second approach is complementary to indexing and it is focused on preprocessing of the (nonlinear) tree pattern and creation of a matching algorithm. The results of the dissertation thesis are divided into two parts. The first, indexing, approach is covered by two different tree indexes. The second, matching, approach is covered by a single tree pattern matching algorithm designed for various tree representations. The first approach is represented by a nonlinear tree pattern pushdown automaton, which can be used to locate occurrences of (nonlinear) tree patterns and a full and linear index also capable of locating occurrences of tree patterns and in extended variant also of nonlinear tree patterns. The second approach is represented by a backward linearised tree pattern matching algorithm, which is a variant on backward pattern matching algorithm known from the area of strings. The algorithm is designed to work with many linear representations of trees. An extension of this algorithm for nonlinear tree patterns is also presented. Tree pattern is a representation of a subgraph of a tree, which is rooted in some node of the tree and contains a wildcard symbol in leaves representing any subtree. The nonlinear tree pattern additionally contains nonlinear variables in leaves which represent any subtree again, however, the same nonlinear variables represent the same subtrees. Given a tree with n nodes, the number of distinct tree patterns and nonlinear tree patterns can be at most 2n−1 + n − 1 and at most (2 + v)n−1 + n − 1, respectively, where v is the number of nonlinear variables allowed in the nonlinear tree patterns.; Trees are one of the fundamental data structures used in Computer Science. The dissertation thesis contributions are best categorised as a part of arbology research [52]. Arbology research is a counterpart of stringology research. Arbology research deals with trees represented in some linear notations, i.e. like strings with additional properties that encode the tree structure. Many algorithms belonging to the stringology maybe, with some care, adapted to handle trees represented as strings using some linear notation. This dissertation thesis is focused on finding all occurrences of tree patterns and nonlinear tree patterns inside a subject tree. Two different general approaches of solving the problem are explored in the dissertation thesis. The first approach is focused on preprocessing of the subject tree and forming a complete index of the subject tree capable of reporting the occurrences when queried with (nonlinear) tree patterns. The second approach is complementary to indexing and it is focused on preprocessing of the (nonlinear) tree pattern and creation of a matching algorithm. The results of the dissertation thesis are divided into two parts. The first, indexing, approach is covered by two different tree indexes. The second, matching, approach is covered by a single tree pattern matching algorithm designed for various tree representations. The first approach is represented by a nonlinear tree pattern pushdown automaton, which can be used to locate occurrences of (nonlinear) tree patterns and a full and linear index also capable of locating occurrences of tree patterns and in extended variant also of nonlinear tree patterns. The second approach is represented by a backward linearised tree pattern matching algorithm, which is a variant on backward pattern matching algorithm known from the area of strings. The algorithm is designed to work with many linear representations of trees. An extension of this algorithm for nonlinear tree patterns is also presented. Tree pattern is a representation of a subgraph of a tree, which is rooted in some node of the tree and contains a wildcard symbol in leaves representing any subtree. The nonlinear tree pattern additionally contains nonlinear variables in leaves which represent any subtree again, however, the same nonlinear variables represent the same subtrees. Given a tree with n nodes, the number of distinct tree patterns and nonlinear tree patterns can be at most 2n−1 + n − 1 and at most (2 + v)n−1 + n − 1, respectively, where v is the number of nonlinear variables allowed in the nonlinear tree patterns.
Fri, 05 Apr 2019 00:00:00 GMThttp://hdl.handle.net/10467/818242019-04-05T00:00:00ZManipulace kapacitou doporučovacích modelů pro optimalizaci v Recall-Coverage rovině
http://hdl.handle.net/10467/81823
Manipulace kapacitou doporučovacích modelů pro optimalizaci v Recall-Coverage rovině; Manipulating the Capacity of Recommendation Models in Recall-Coverage Optimization
Tomáš Řehořek
Traditional approaches in Recommender Systems ignore the problem of long-tail recommendations. There is no systematic approach to control the magnitude of long-tail recommendations generated by the models, and there is not even proper methodology to evaluate the quality of long-tail recommendations. This thesis addresses the long-tail recommendation problem from both the algorithmic and evaluation perspective. We proposed controlling the magnitude of long-tail recommendations generated by models through the manipulation with capacity hyperparameters of learning algorithms, and we dene such hyperparameters for multiple state-of-the-art algorithms. We also summarize multiple such algorithms under the common framework of the score function, which allows us to apply popularity-based regularization to all of them. We propose searching for Pareto-optimal states in the Recall-Coverage plane as the right way to search for long-tail, high-accuracy models. On the set of exhaustive experiments, we empirically demonstrate the corectness of our theory on a mixture of public and industrial datasets for 5 dierent algorithms and their dierent versions.; Traditional approaches in Recommender Systems ignore the problem of long-tail recommendations. There is no systematic approach to control the magnitude of long-tail recommendations generated by the models, and there is not even proper methodology to evaluate the quality of long-tail recommendations. This thesis addresses the long-tail recommendation problem from both the algorithmic and evaluation perspective. We proposed controlling the magnitude of long-tail recommendations generated by models through the manipulation with capacity hyperparameters of learning algorithms, and we dene such hyperparameters for multiple state-of-the-art algorithms. We also summarize multiple such algorithms under the common framework of the score function, which allows us to apply popularity-based regularization to all of them. We propose searching for Pareto-optimal states in the Recall-Coverage plane as the right way to search for long-tail, high-accuracy models. On the set of exhaustive experiments, we empirically demonstrate the corectness of our theory on a mixture of public and industrial datasets for 5 dierent algorithms and their dierent versions.
Thu, 04 Apr 2019 00:00:00 GMThttp://hdl.handle.net/10467/818232019-04-04T00:00:00Z