Jiri Stastny

Comparison of learning algorithms

There are several learning methods which are suitable for neural networks. In this paper two of them are described - Back-propagation (BP) and Genetic (GA) algorithms. These learning methods are compared here and they are used for the control of modern telecommunication network nodes.

Analysis of Algorithms for Radial Basis Function Neural Network

This paper describes the analysis of algorithms for the hidden layer construction of network and for learning of the Radial Basis Function neural Network (RBFN). We compared results obtained by using of learning algorithms LMS (Least Mean Square) and Gradient Algorithms (GA) and results are obtained by using of algorithms APC-III and K-means for hidden layer contruction of neural network. The principles and algorithms given below have been used in an application for object classification that was developed at Brno University of Technology. This solution is suitable for the research of personal wireless communications and similar systems.

Wavelet transform for image analysis

The wavelet transform is a comparatively new and fast developing method for analysing signals. The main advantage of applying the wavelet transform to the detection of edges in an image is the possibility of choosing the size of the details that are detected. The size of detected edges is set by the wavelet scale. In the case of the discrete wavelet transform the choice of the scale is performed by multiple signal passage through the wavelet filter. When processing a 2-D image, the wavelet analysis is performed separately for the horizontal and the vertical function. The vertical and the horizontal edges are thus detected separately. The wavelet transform splits the input signal into two components. One contains the low-frequency (LP) part of input signal, which corresponds to major changes in the function (individual objects in the image, etc.). The other part contains the high-frequency (HP) part of input signal, which corresponds to details in the function (noise, edges, etc.). This signal component is not processed on the next level of transformation.

Ensuring invariances for structural methods of object recognition

The paper discusses ensuring invariances for the structural methods of recognition of randomly deformed object. Initially, the types of invariances are described. Further, the ways of effective ensuring all types of invariances are described, mainly using automatic selection of point of description origin, differential primitives and object rotation. Finally, the results of this method are evaluated.

Visualization of uncertainty in LANDSAT classification process

Many uncertainties can be found in the classification of remotely sensed data. Namely they can arise in defining classification classes. We use two ways, incorporating acquired Corine Land Cover labels and our manually annotated labels. We are describing several visualization possibilities to demonstrate uncertainties in labels and their connections with classification results. We use parallel coordinates to visualize data, presenting problems in classes definitions. These failures can be consequently seen in the results of classification in confusion matrix. We inspect also posterior probabilities of k-Nearest Neighbor (k-NN) classifier visualizing maximum likelihood class probabilities as alpha channel of resulting classification map.

Audio data classification by means of new algorithms

This paper describes classification of sound recordings based on their audio features. This is useful for querying large datasets, searching for recordings with some desired content. We use musical recordings as well as birdsongs recordings, which usually have rich structure and contain a lot of patterns suitable for classification. We present two different classification methods, one for musical recordings and one for birdsongs. These methods are compared and their differences are discussed. We use feature vectors that capture the audio content of recording as a whole piece and then classify these feature vectors using combination of the Self-organizing map and the Learning Vector Quantization, which represent a powerful algorithm using unlabeled as well as labeled data. In case of birdsongs we use feature vectors representing time frames of a recording.

Artificial Neural Networks Numerical Forecasting of Economic Time Series

The current global market is driven by many factors, e.g. by the facts that we live in the information age and that information is distributed in short times, large amounts and by many data channels. It is practically impossible to analyse all kinds of incoming information flows and transform them to data by classical methods. New requirements call for new methods. Artificial neural networks once trained on patterns can be used for forecasting and they are able to work with extremely big datasets in reasonable time. Traditionally, this is solved by means of a statistical analysis first a time-series model is constructed and then statistical prediction algorithms are applied to it in order to obtain future values. The common point for both methods is the learning process from samples of past data, or learning from the past. From many of the uncommon points the input conditions for the model creation and the length of the time series pattern set could be pointed out. On one hand, very sophisticated statistical methods exist that have strictly defined input conditions for datasets; on the other hand, practically open input conditions of artificial neural networks can be used. Regarding the length of the time series, the main problem of the Czech Republic, short and middle term predictions are valuable datasets. The lengths of selected economic values are not huge enough for quality of prediction or forecasting. Hand-in-hand with typical problems of real datasets (noisiness and/or missing data), there is the issue of the quality of the numerical forecasting. In addition, the strong nonlinearity of the models leads to an unsolvable usage of classical methods or construction of models that are not representing the reality. These are only few of the difficulties related to economic and financial modelling and prediction. Possible problems of numerous types of the artificial neural networks with n-setups make the issue even more complicated. The aim of this chapter is to compare different types of artificial neural networks using short and middle terms predictions of a real-world economic index. A number of papers dealing with artificial neural networks used for particular problems and often for the test do not use real-world economic indexes. The chapter is divided into four sections. The first simply presents the introduction to the research domain. The second section describes state-of-the-art artificial intelligence approaches to both prediction and forecasting of economic indexes. In the third section, neural network types and learning algorithms dealing with the prediction of time series and learning optimization are presented. In detail, the third section also includes methods of verification and validation of artificial neural networks and description of real-world economic indexes

Object recognition by means of early parser effective implementation

The paper discusses effective implementation of Earley parser for structural methods of recognition of randomly deformed object. Initially, basics of Earley parser are described. Further, ways of effective implementation and improvements of this method are described, mainly using prediction look-ahead and grammar optimization for improvement of the analysis. Finally, results of this method are evaluated.

Traveling Salesman Problem optimization by means of graph-based algorithm

There are many different algorithms for optimization of logistic and scheduling problems and one of the most known is Genetic algorithm. In this paper we take a deeper look at a draft of new graph-based algorithm for optimization of scheduling problems based on Generalized Lifelong Planning A* algorithm which is usually used for path planning of mobile robots. And then we test it on Traveling Salesman Problem (TSP) against classic implementation of genetic algorithm. The results of these tests are then compared according to the time of finding the best path, its travel distance, an average distance of travel paths found and average time of finding these paths. A comparison of the results shows that the proposed algorithm has very fast convergence rate towards an optimal solution. Thanks to this it reaches not only better solutions than genetic algorithm, but in many instances it also reaches them faster.