TAAC'2012, Section Three: Artificial Intelligence

The paper present results of the Ukrainian-French scientific activity between Donetsk National Technical University and French partners in the field of human-robot and robot-environment interactions. The first part of the research shows adaptation of robotic arm movements to the dynamics of interacting person who imposes its own rhythmicity. The adaptation algorithm is based on rhythmic oscillator inspired by biology. The second part of the research considers adaptation of a biped robot internal vibration modes to the environment during flexion-extension vertical movements. These vibrations are extremely unwanted; they appear during robot motion by reason of internal flexibilities in articulations, backlashes and friction and can cause the fall of the robot. It is shown that their efficient attenuation can be achieved by using auto-adaptive oscillator that acts on the robot's knee articulations.

Architecture and algorithms for ensemble of adaptive fuzzy Kohonen clustering neural networks are considered. The networks of the ensemble operate asynchronously in real-time with different assumptions about character of receiving data. Usage of possibilistic and robust algorithms cause to steady results in terms of outliers in data and faults in adjustment, in combination with flexible adaptation to non-stationary data phenomenon.

The main aim of this work, was to develop system, based on artificial neural networks, which can recognize handwritten symbols. It also can be used for gestures recognition. Due to the significant development communicators, smartphones, etc., this problem is extremely urgent. Recognition rates for the system vary depending on the consistency of the writing. On average, the untrained system achieved 70\% recognition. After training, average recognition rates of 90\% were achieved

The paper displays main results of the Ukrainian-French research project which studies interaction of a robot with its environment. Following applications of computer vision technology for various issues of robotics are shown: robot follows movements of a person, detects speed of the moving object, studies its environment, learns particular manipulations, gets trained to focus attention on the objects of interest and synchronizes its motion with an interacting agent. The article proposes approaches to development of the software interface for the robot control based on the computer vision technology.

This paper describes the most common approaches to sentiment analysis and defines an optimal approach for deep sentiment analysis of restaurant reviews in Ukrainian.

The problem of distorted digital images recognition using a stripe classification algorithm based on pseudoinverse methods is analysed. The technique of digital image recognition and its computer implementation in the software environment engineering and mathematical package Matlab is suggested. This method is tested on numbers and letters of the Ukrainian alphabet. The comparative characteristic of stripe method of classification, method of neural networks, and support vector method was made.

Determination of corner point position in subpixel accuracy has a high significance in many practical applications. This paper deals with subpixel accuracy of Harris corner detector and comparison of $2$ approaches to specify the results. For our experiment we used very accurate pointing device and there was set of images for every position taken. The main goal is to use several images in static scene for getting more accurate and more robust results. First approach is based on averaging of found corner point coordinates among all images in one set. The second one is about averaging of brightness intensity for all images in one set before we apply the detector. We compared both these approaches and there were appropriate statistical analysis performed. All the results are illustrated in graphs and listed in tables. This comparison and our study could be convenient in many types of applications and measurements in case the precision is a key.

Development questions of gathering, storage and data processing systems with a considerable quantity of the sensors metering various environmental factors which can represent danger to human health and biosphere are considered. The construction of a system prototype which includes directly the supersensory data collection system and client-server architecture is offered. Also questions of the supersensory networks construction and a complex data analysis are considered.

We consider a class of pursuit-evasion-games with incomplete information-exchange modeled on a real data. Pursuer and evader agents are acting on geographical maps having incomplete (noised) information about each others strategies. We describe the optimal strategies for several information-patterns using Kalman filter approach to eliminate noise related to information corruption. Linear form solutions are given based on a geometrical approach. Then attention is focused on a game in which some of the opponent's state variables and his strategy are known. The player-system is approximated by another system to which observer-theory can be applied. The practical application of the methods stated above is shown in comparison to other methods used for solving similar problems. Generalization of the Kalman filters are discussed in application to N-agent systems and other types of information exchange. Software product using real geolocation data is presented.