Rudolf Jaksa

Stochastic weight update in the backpropagation algorithm on feed-forward neural networks

We will examine stochastic weight update in the backpropagation algorithm on feed-forward neural networks. It was introduced by Salvetti and Wilamowski in 1994 in order to improve probability of convergence and speed of convergence. However, this update method has also one another quality, its implementation is simple for arbitrary network topology. In stochastic weight update scenario, constant number of weights is randomly selected and updated. This is in contrast to classical ordered update, where always all weights are updated. We will describe exact implementation, and present example results on toy-task data with feed-forward neural network topology. Stochastic weight update is suitable to replace classical ordered update without any penalty on implementation complexity and with good chance without penalty on quality of convergence.

Tuning of image parameters by interactive evolutionary computation

We discuss the use of interactive evolutionary computation for designing and optimizing image enhancement filters with fewer parameters than normal filter model. We present a method to speed up this optimization by combining interactive and non-interactive evaluation.

Clustering of users inputs in multi-user interactive evolutionary font design

Clustering of users inputs in multi-user Interactive Evolutionary Computation is intended to allow to collect large data sets for user-behavior modeling, while preserving the users individuality. By clustering the user input data, we group together similar behaviors and distribute opposite ones, thus preventing conflicts in data resulting from opposite opinions of different users. Without this clustering, it might be not possible to use data obtained from several users for the behavior modeling. This paper tries to present the application of Self-Organizing-Map clustering in the task of font design with Interactive Evolutionary Computation interface.

Local weather prediction system for a heating plant using cognitive approaches

Graphical abstractDisplay Omitted HighlightsA structure of a weather prediction system is proposed.Realized measuring points for data collection and transfer of weather variables.Designed and tested prediction model based on chained neural networks.Realized means for data modification based on fuzzy logic.A prediction system is implemented in a heating plant. Present-day requirements emphasize the need of saving energy. It relates mainly to industrial companies, where the minimization of energy consumption is one of their most important tasks they face. In our paper, we deal with the design of the so-called weather prediction system (WPS) for the needs of a heating plant. The primary task of such a WPS is timely predicting expected heat consumption to prepare the technology characterized by long delays in advance. Heat prediction depends primarily on weather so the crucial part of WPS is the weather, especially temperature, prediction. However, a prediction system needs a variety of further data, too. Therefore, WPS must be regarded as a complex system, including data collection, its processing, own prediction and eventual decision support. This paper gives the overview about existing data processing systems and prediction methods and then it describes a concrete design of a WPS with distributed data measuring points (stations), which are processed using a structure of neural networks based on multilayer perceptrons (MLP) with a combination of fuzzy logic. Based on real experiments we show that also such simple means as MLPs are able to solve complex problems. The paper contains a basic methodology for designing similar WPS, too.

Neural Networks Based Local Weather Prediction System

In this paper we describe how to build a fully autonomous system for collection, prediction and presentation of single-position meteorological data the local weather prediction system. By employing nonlinear statistics with neural network predictor on meteorological time-series data we were able to achieve good results for the one-day weather prediction. This novel local statistical approach to weather prediction is different compare to standard methods which are based on the air mass movement modelling. Main objective of this paper is to describe whole system for local weather prediction including technology, software, methods and parameters, and also experimental results.

Daily Temperature Profile Prediction for the District Heating Application

We show an application of artificial neural networks for local weather prediction. By employment of appropriate network structure and proper selection of input/output signals, solid results can be achieved. Our system was implemented in the local district heating company, where it was used to predict daily temperature profile with period of 15 minutes. Further, weekly and yearly profiles were predicted, and also heat consumption profiles. Whole prediction system consists of several chained neural networks and data processing modules. Training data for neural networks were collected from meteorological stations around the Kosice city. Additional training data were collected by web-robots from internet from several weather forecast agencies.

Framework for the interactive learning of artificial neural networks

We propose framework for interactive learning of artificial neural networks. In this paper we study interaction during training of visualizable supervised tasks. If activity of hidden node in network is visualized similar way as are network outputs, human observer might deduce the effect of this particular node on the resulting output. We allow human to interfere with the learning process of network, thus he or she can improve the learning performance by incorporating his or her lifelong experience. This interaction is similar to the process of teaching children, where teacher observes their responses to questions and guides the process of learning. Several methods of interaction with neural network training are described and demonstrated in the paper.

Stochastic weight update for recurrent networks

Stochastic weight update is a variant of error back-propagation algorithm for learning of artificial neural networks. It allows for efficient topology-independent implementation of backpropagation through time for recurrent networks. In stochastic weight update scenario, constant number of weights and neurons is randomly selected and updated. This is in contrast to the classical ordered update, where all weights/neurons are always updated. In this paper we will study performance of stochastic weight update on recurrent neural networks using concept of feedforward network with added recurrent neurons.

Influence of Number of Neurons in Time Delay Recurrent Networks with Stochastic Weight Update on Backpropagation Through Time

We will examine the various modifications of backpropagation through time algorithm (BPTT) done by stochastic update in the recurrent neural networks (RCNN) including the influence of the different numbers of recurrent neurons. The general introduction involving the stochasticity into neural network was provided by Salvetti and Wilamowski in 1994 in order to improve probability of convergence and speed of convergence. The implementation is simple for arbitrary network topology. In stochastic update scenario, constant number of weights and neurons (neurons selected before starting learning phase) are randomly selected and updated. This is in contrast to classical ordered update, where always all weights or neurons are updated. Stochastic update is suitable to replace classical ordered update without any penalty on implementation complexity and with good chance without penalty on quality of convergence. We have provided first experiments with stochastic modification on backpropagation algorithm (BP) used for artificial feed-forward neural network (FFNN) in detail described in our paper [1]. We will present experiment results on simple toy-task data of time shifted and skewed signal as a verification of our implementation of different algorithm modifications.

Robotic attention manager using fuzzy controller with fractal analysis

This paper is focused on the application of fractal analysis in the attention management of humanoid robot. We designed a fuzzy controller to combine the face detection, movement detection and the fractal dimension signals to control the head movement of robot Nao. Also, the gaze problem is addressed by the controller. Implementation details are included in the paper, including configuration parameters, which we found optimal according to subjective analysis and possibilities of current hardware. We found the fuzzy controller to be advantageous for implementation of attention manager because of smoothing of the movement of robot when compared to the simple rule based implementation, and also because the fuzzy controller implementation of manager is more clear than a naive if-then heuristics code. We also found the fractal dimension to be useful additional signal for attention management of robot, which can be computed in near real-time on current hardware and static input images.