Michal Janosek

Knowledge discovery in dynamic data using neural networks

The paper proposes a new approach to implement common neural network algorithms in the network environment. In our experimental study we have used three different types of neural networks based on Hebb, daline and backpropagation training rules. Our goal was to discover important market (Forex) patterns which repeatedly appear in the market history. Developed classifiers based upon neural networks should effectively look for the key characteristics of the patterns in dynamic data. We focus on reliability of recognition made by the described algorithms with optimized training patterns based on the reduction of the calculation costs. To interpret the data from the analysis we created a basic trading system and trade all recommendations provided by the neural network.

Cryptography Based On Neural Network.

The goal of cryptography is to make it impossible to take a cipher and reproduce the original plain text without the corresponding key. With good cryptography, your messages are encrypted in such a way that brute force attacks against the algorithm or the key are all but impossible. Good cryptography gets its security by using incredibly long keys and using encryption algorithms that are resistant to other form attack. The neural net application represents a way of the next development in good cryptography. This paper deals with using neural network in cryptography, e.g. designing such neural network that would be practically used in the area of cryptography. This paper also includes an experimental demonstration. INTRODUCTION TO CRYPTOGRAPHY The cryptography deals with building such systems of security of news that secure any from reading of trespasser. Systems of data privacy are called the cipher systems. The file of rules are made for encryption of every news is called the cipher key. Encryption is a process, in which we transform the open text, e.g. message to cipher text according to rules. Cryptanalysis of the news is the inverse process, in which the receiver of the cipher transforms it to the original text. The cipher key must have several heavy attributes. The best one is the singularity of encryption and cryptanalysis. The open text is usually composed of international alphabet characters, digits and punctuation marks. The cipher text has the same composition as the open text. Very often we find only characters of international alphabet or only digits. The reason for it is the easier transport per media. The next cipher systems are the matter of the historical sequence: transposition ciphers, substitution ciphers, cipher tables and codes. Simultaneously with secrecy of information the tendency for reading the cipher news without knowing the cipher key was evolved. Cipher keys were watched very closely. The main goal of cryptology is to guess the cipher news and to reconstruct the used keys with the help of good analysis of cipher news. It makes use of mathematical statistics, algebra, mathematical linguistics, etc., as well as known mistakes made by ciphers too. The legality of the open text and the applied cipher key are reflected in every cipher system. Improving the cipher key helps to decrease this legality. The safety of the cipher system lies in its immunity against the decipher. The goal of cryptanalysis is to make it possible to take a cipher text and reproduce the original plain text without the corresponding key. Two major techniques used in encryption are symmetric and asymmetric encryption. In symmetric encryption, two parties share a single encryption-decryption key (Khaled, Noaman, Jalab 2005). The sender encrypts the original message (P), which is referred to as plain text, using a key (K) to generate apparently random nonsense, referred to as cipher text (C), i.e.: C = Encrypt (K,P) (1) Once the cipher text is produced, it may be transmitted. Upon receipt, the cipher text can be transformed back to the original plain text by using a decryption algorithm and the same key that was used for encryption, which can be expressed as follows: P = Dencrypt (K,C) (2) In asymmetric encryption, two keys are used, one key for encryption and another key for decryption. The length of cryptographic key is almost always measured in bits. The more bits that a particular cryptographic algorithm allows in the key, the more keys are possible and the more secure the algorithm becomes. The following key size recommendations should be considered when reviewing protection (Ferguson, Schneier, Kohno, 2010): Symmetric key: • Key sizes of 128 bits (standard for SSL) are sufficient for most applications Proceedings 26th European Conference on Modelling and Simulation ©ECMS Klaus G. Troitzsch, Michael Mohring, Ulf Lotzmann (Editors) ISBN: 978-0-9564944-4-3 / ISBN: 978-0-9564944-5-0 (CD) • Consider 168 or 256 bits for secure systems such as large financial transactions Asymmetric key: • Key sizes of 1280 bits are sufficient for most personal applications • 1536 bits should be acceptable today for most secure applications • 2048 bits should be considered for highly protected applications. Hashes: • Hash sizes of 128 bits (standard for SSL) are sufficient for most applications • Consider 168 or 256 bits for secure systems, as many hash functions are currently being revised (see above). NIST and other standards bodies will provide up to date guidance on suggested key sizes. BACKPROPAGATION NEURAL NETWORKS An Artificial Neural Network (ANN) is an information processing paradigm that is inspired by the way biological nervous systems, such as the brain, process information. The key element of this paradigm is the structure of the information processing system. It is composed of a large number of highly interconnected processing elements (neurones) working in unison to solve specific problems. ANNs, like people, learn by example. An ANN is configured for a specific application, such as pattern recognition or data classification, through a learning process. Learning in biological systems involves adjustments to the synaptic connections that exist between the neurones. This is true of ANNs as well. Figures 1: A general three layer neural network Backpropagation network is one of the most complex neural networks for supervised learning. Regarding topology, the network belongs to a multilayer feedforward neural network. See Fig. 1 (Volna 2000), usually a fully connected variant is used, so that each neuron from the n-th layer is connected to all neurons in the (n+1)-th layer, but it is not necessary and in general some connections may be missing – see dashed lines, however, there are no connections between neurons of the same layer. A subset of input units has no input connections from other units; their states are fixed by the problem. Another subset of units is designated as output units; their states are considered the result of the computation. Units that are neither input nor output are known as hidden units. Figures 2: A simple artificial neuron (http://encefalus.com/neurology-biology/neuralnetworks-real-neurons) A basic computational element is often called a neuron (Fig. 2), node or unit (Fausett 1994). It receives input from some other units, or perhaps from an external source. Each input has an associated weight w, which can be modified so as to model synaptic learning. The unit computes some function f of the weighted sum of its inputs (3):

Pattern Recognition Algorithm Optimization

In this article, a short introduction into the field of pattern recognition in time series has been given. Our goal is to find and recognize important patterns which repeatedly appear in the market history. We focus on reliability of recognition made by the proposed algorithms with optimized patterns based on artificial neural networks. The performed experimental study confirmed that for the given class of tasks can be acceptable a simple Hebb classifier with a proposed modification that has been designed, tested, and used for the active mode of Hebb rule. Finally, we present comparison results of trading based on both recommendations: using proposed Hebb neural network implementation, and human expert.

Methodology for System Adaptation Based on Characteristic Patterns

This paper describes the methodology for system description and application so that the system can be managed using real time system adaptation. The term system here can represent any structure regardless its size or complexity (industrial robots, mobile robot navigation, stock market, systems of production, control systems, etc.). The methodology describes the whole development process from system requirements to software tool that will be able to execute a specific system adaptation. In this work, we propose approaches relying on machine learning methods (Bishop, 2006), which would enable to characterize key patterns and detect them in real time and in their altered form as well. Then, based on the pattern recognized, it is possible to apply a suitable intervention to system inputs so that the system responds in the desired way. Our aim is to develop and apply a hybrid approach based on machine learning methods, particularly based on soft-computing methods to identify patterns successfully and for the subsequent adaptation of the system. The main goal of the paper is to recognize important pattern and adapt the system’s behaviour based on the pattern desired way. The paper is arranged as follows: Section 1 introduces the critical topic of the article. Section 2 details the feature extraction process in order to optimize the patterns used as inputs into experiments. The pattern recognition algorithms using machine learning methods are discussed in section 3. Section 4 describes the used data-sets and covers the experimental results and a conclusion is given in section 5. We focus on reliability of recognition made by the described algorithms with optimized patterns based on the reduction of the calculation costs. All results are compared mutually.

Nonlinear Time Series Analysis via Neural Networks

This article deals with a time series analysis based on neural networks in order to make an effective forex market [Moore and Roche, J. Int. Econ. 58, 387–411 (2002)] pattern recognition. Our goal is to find and recognize important patterns which repeatedly appear in the market history to adapt our trading system behaviour based on them.

Simulation Parameters Settings Methodology Proposal Based on Leverage Points

Simulation belongs to one of the most time consuming phases of complex system design. It is necessary to test our model with number of parameters of the entire simulation, control mechanism and components. We test various scenarios and strategies. In this article we would like to present the methodology proposal for simulation parameters settings based on leverage points’ hierarchy developed by Donella H. Meadows to aid the simulation process.

Predator-Prey Simulation's Parameters And Leverage Points.

For many natural systems there is an advantage in their simulating on a computer. In this article, a way how to determine the influence of the predator-prey simulation’s parameters on the model’s behaviour based on the leverage points is examined. An investigation whether it is possible to sort the parameters according to the importance and if it is possible to automate the process to a certain degree is performed.

Pattern Recognition and Classification in Time Series Data

Patterns can be any number of items that occur repeatedly, whether in the behaviour of animals, humans, traffic, or even in the appearance of a design. As technologies continue to advance, recognizing, mimicking, and responding to all types of patterns becomes more precise. Pattern Recognition and Classification in Time Series Data focuses on intelligent methods and techniques for recognizing and storing dynamic patterns. Emphasizing topics related to artificial intelligence, pattern management, and algorithm development, in addition to practical examples and applications, this publication is an essential reference source for graduate students, researchers, and professionals in a variety of computer-related disciplines.

Machine learning approach to point localization system

The article introduces point localization systems in 3D Euclidean space based on neural networks. There are two models presented. The first one identified distances between a randomly generated point and a reference points in the defined domain. Then a neural network uses the obtained distances as its inputs to determine the actual position of the point in the domain space. Due to a relatively good accuracy that was obtained during the experimental study, the proposed model based on neural networks was used in the second model as an acoustic Motion Capturing system (MoCap). MoCap system is represented by a neural network that uses obtained distances between transmitters and a receiver as its inputs to determine an actual position of the receiver in space. We also propose a new way to minimize a training set by using ANFIS approach in this specific problem. All obtained results are summarized in the conclusion.

Preliminary multivariate analysis of the Harvard spectral classification of the H-R diagram main sequence stars

This article aims to analyse main sequence stars of the Hertzsprung–Russell diagram using cluster analysis methods in order to prove if the results from the cluster analysis will match with Harvard spectral classification. Nextdo a dimensional reduction task in order to see if it is possible to explain the variance of the data using less factors than examined numbers of parameters. Using particular stars’ parameters following methods are performed: hierarchical clustering, centroid-based clustering (k-means), principal component analysis and factor analysis.