|
|
Learning with Regularization Networks
Kudová, Petra ; Neruda, Roman (advisor) ; Andrejková, Gabriela (referee) ; Hlaváčková-Schindler, Kateřina (referee)
In this work we study and develop learning algorithms for networks based on regularization theory. In particular, we focus on learning possibilities for a family of regularization networks and radial basis function networks (RBF networks). The framework above the basic algorithm derived from theory is designed. It includes an estimation of a regularization parameter and a kernel function by minimization of cross-validation error. Two composite types of kernel functions are proposed - a sum kernel and a product kernel - in order to deal with heterogenous or large data. Three learning approaches for the RBF networks - the gradient learning, three-step learning, and genetic learning - are discussed. Based on the se, two hybrid approaches are proposed - the four-step learning and the hybrid genetic learning. All learning algorithms for the regularization networks and the RBF networks are studied experimentally and thoroughly compared. We claim that the regularization networks and the RBF networks are comparable in terms of generalization error, but they differ with respect to their model complexity. The regularization network approach usually leads to solutions with higher number of base units, thus, the RBF networks can be used as a 'cheaper' alternative in terms of model size and learning time.
|
|
|
Modely výpočetní inteligence pro hydrologické predikce
Paščenko, Petr ; Neruda, Roman (advisor) ; Petříčková, Zuzana (referee)
The thesis deals with the application of computational artificial intelligence models on hydrological predictions. The short term rainfall-runoff prediction problem is studied on the real data of physical time seriesmeasured in the watershed of river Plučnice. A brief statistical study including correlation and regression analyses is performed. The high level of variance and noise is concluded. The evolution of the proper input filter providing an input set for the neural network is performed. In the main part of the thesis several neural network models based on multilayer perceptron, RBF units, and neuroevoution are constructed together with two neural ensembles inspired by the bagging method. The models are tested on the three subsequent years summer data. The greater generalization ability of multilayer perceptron architectures is concluded. The resulting multilayer perceptron models are able to reduce the mean squared error of the prediction by 15% compared to the prediction by the previous value.
|
|
|
Parallel evolutionary algorithms for multiobjective optimization
Pilát, Martin ; Neruda, Roman (advisor) ; Mráz, František (referee)
In the present work we study the options for parallelization of evolutionary algorithms for multiobjective optimization (MOGA). We provide the overview of existing sequential and parallel MOGAs and we propose three other methods: FCMOGA - MOGA with fuzzy constraints, HIMOGA - heterogeneous island MOGA, and MOGASOLS - MOGA with single objective local search. We test these algorithms on a set of benchmark problems and compare them with existing MOGAs.
|
|
|
EASIMEN - Sandbox pro umělé bytosti v simulovaném prostředí
Hencz, Attila ; Neruda, Roman (advisor) ; Pilát, Martin (referee)
Title: EASIMEN - A sandbox for artificial creatures in simulated envi- ronment Author: Attila Hencz Department / Institute: Department of Theoretical Computer Sci- ence and Mathematical Logic Supervisor of the bachelor thesis: Mgr. Roman Neruda, CSc., Insti- tute of Computer Science of the ASCR,1 v.v.i.2 Supervisor's e-mail address: Roman.Neruda@mff.cuni.cz Abstract: The present work investigates the creation of a flexible near- universal simulation environment for artificial intelligence with embodi- ment and real-time simulation in mind. Also, a look is taken at the prob- lems associated with the functioning of such a systems and the methods that can be used in a simulation of this kind. The current implementation of the simulation environment (EASIMEN) is rather simplified and its purpose is the demonstration of the proposed design and architecture. Additionally, there are a couple of simplistic module implementations available for the underlying artificial intelligence architecture proposed by the author (BIAR). These modules are only mere showcases of certain aspects of the issues at hand, and serve as templates for the implemen- tation of more sophisticated modules in the future. Keywords: artificial intelligence, real-time simulation, 3D graphics, phys- ically simulated virtual environment, embodied...
|
|
|
Multi-Agent systems and organizations
Kúdela, Lukáš ; Štěpánek, Petr (advisor) ; Neruda, Roman (referee)
Multi-agent systems (MAS) are emerging as a promising paradigm for conceptualizing, designing and implementing large-scale heterogeneous software systems. The key advantage of looking at components in such systems as autonomous agents is that as agents they are capable of flexible self-organization, instead of being rigidly organized by the system's architect. However, self-organization is like evolution-it takes a lot of time and the results are not guaranteed. More often than not, the system's architect has an idea about how the agents should organize themselves-what types of organizations they should form. In our work, we tried to solve the problem of modelling organizations and their roles in a MAS, independent of the particular agent platform on which the MAS will eventually run. First and foremost, we have proposed a metamodel for expressing platform-independent organization models. Furthermore, we have implemented the proposed metamodel for the Jade agent platform as a module extending this framework. Finally, we have demonstrated the use of our module by modelling three specific organizations: remote function invocation, arithmetic expression evaluation and sealed-bid auction. Our work shows how to separate the behaviour acquired through a role from the behaviour intrinsic to an agent. This...
|
|
|
Evoluční algoritmy pro strukturální učení neuronových sítí
Kasík, Pavel ; Neruda, Roman (advisor) ; Kudová, Petra (referee)
Designing neural networks topologies is s complicated problem when we consider general network structures. Evolutionary algorithm can provide us with interesting solutions of this problem. This work introduces an evolutionary algorithm for evolving neural networks. One of the possible algorithms for evolving neural networks is the NEAT algorithm. The goal of this work is to modify and enhance abilities of the NEAT algorithm. Improvements are focused on utilizing position of a neuron in network, improving crossover procedure and introducing solution of algorithm parallelization that preserve abilities of both NEAT and the new algorithm.
|
|
|
EASIMEN - Sandbox pro umělé bytosti v simulovaném prostředí
Hencz, Attila ; Neruda, Roman (advisor) ; Brom, Cyril (referee)
In the present work we investigate the creation of a exible nearuniversal simulation environment for arti cial intelligence with embodiment and real-time simulation in mind. We also take a look at the problematics associated with the functioning of such an systems and the methods that can be used in a simulation of this kind. The current implementation of the simulation environment is not a complete system, but rather more separate projects which are showcases of certain aspects of the problematics at hand.
|
|
|
Evolutionary Algorithms for Multiobjective Optimization
Pilát, Martin ; Neruda, Roman (advisor) ; Schoenauer, Marc (referee) ; Pošík, Petr (referee)
Multi-objective evolutionary algorithms have gained a lot of atten- tion in the recent years. They have proven to be among the best multi-objective optimizers and have been used in many industrial ap- plications. However, their usability is hindered by the large number of evaluations of the objective functions they require. These can be expensive when solving practical tasks. In order to reduce the num- ber of objective function evaluations, surrogate models can be used. These are a simple and fast approximations of the real objectives. In this work we present the results of research made between the years 2009 and 2013. We present a multi-objective evolutionary algo- rithm with aggregate surrogate model, its newer version, which also uses a surrogate model for the pre-selection of individuals. In the next part we discuss the problem of selection of a particular type of model. We show which characteristics of the various models are im- portant and desirable and provide a framework which combines sur- rogate modeling with meta-learning. Finally, in the last part, we ap- ply multi-objective optimization to the problem of hyper-parameters tuning. We show that additional objectives can make finding of good parameters for classifiers faster. 1
|
|
|
Universality in Amorphous Computing
Petrů, Lukáš ; Wiedermann, Jiří (advisor) ; Janeček, Jan (referee) ; Neruda, Roman (referee)
Amorphous computer is a theoretical computing model consisting of randomly located tiny devices (called nodes) in some target area. The nodes of an amorphous computer can communicate using short-range radio. The communication radius is small compared to the size of the target area. The nodes are all identical, initially have no identi ers, work asynchronously and there is no standard communication protocol. An amorphous computer must work for any number of nodes under reasonable statistical assumptions concerning the spatial distribution of nodes. Moreover, the computation should use very limited amount of memory on each node. For the just described concept of amorphous computer we investigate the question whether a universal computation is possible at all in a corresponding theoretical model. To answer this question, several subsequent steps are performed. In the rst step, we design a formal minimalist model of a node and of the amorphous computer as a whole. In the second step, we develop communication protocol for the amorphous computer. In the last step, we show the universality by simulating a computation of a universal machine. The size of the amorphous computer will depend on the space complexity of the simulated machine. All the previously mentioned steps are described in detail in this work....
|
|
|
Control algorithms for autonomous embodied agents
Slušný, Stanislav ; Neruda, Roman (advisor) ; Kvasnička, Vladimír (referee) ; Koutník, Jan (referee)
Charles University in Prague Faculty of Mathematics and Physics DOCTORAL THESIS Mgr. Stanislav Slušný Control algorithms for autonomous embodied agents Department of Software Engineering Supervisor of the doctoral thesis: Mgr. Roman Neruda, CSc. Study programme: Computer Science Specialization: Software Engineering Prague 2014 Title: Control algorithms for autonomous embodied agents Author: Mgr. Stanislav Slušný Department: Department of Software Engineering, Faculty of Mathematics and Physics, Charles University, Prague Supervisor: Mgr. Roman Neruda, CSc., Institute of Computer Science, Academy of Sciences of the Czech Republic, Prague Abstract: This work studies control algorithms for adaptive embodied agents. The available approaches, based on neural networks, genetic algorithms and re- inforcement learning are investigated and potential improvements suggested. Ar- chitecture of adaptive embodied autonomous agents, that combines the existing reactive and deliberative paradigms, is proposed and demonstrated in a realistic simulator solving a complex real world task. The performance of a novel high-level planner based on constraint programming and finite automata is demonstrated on a practical application.
|
guest ::
