Svitlana Vyshemyrska

Bio: Svitlana Vyshemyrska is an academic researcher from Kherson National Technical University. The author has contributed to research in topics: Interface (computing) & Optimization problem. The author has an hindex of 2, co-authored 8 publications receiving 11 citations.

Dynamic Coordination of Strategies for Multi-agent Systems.

Abstract: Dynamic coordination of multi-agent systems (MAS) strategies under uncertainty based on a stochastic game model is solved. Dynamic coordination is teaching of the system of generating spatially distributed periodic signals. A stochastic game model is built, criteria for dynamic coordination of player strategies are determined, a recurrent method, algorithmic and software for stochastic game solving are developed. The developed model, method and algorithm for stochastic game resolution provide dynamic MAS coordination, which is manifested in locally deter-mined spatial and temporal alignment of players’ strategies. Dynamic coordination is ensured by an adaptive search method for resolving stochastic play, taking into account current penalties for relevant spatial coordination and rhythm disturbances. It is established that the effectiveness of training players to perform coordinated rhythmic actions is determined by the balance between these penalties, which is achieved by the influence of white noise. Dynamic coordination of agent strategies is achieved as the real-time stochastic game is unleashed based on gathering current information and its adaptive processing. The considered method belongs to the class of reactive methods and simulates the reflexive behavior of living organisms. The method allows finding stochastic game solutions in pure strategies.

Some Features of the Numerical Deconvolution of Mixed Molecular Spectra

Abstract: The direct method features of finding the weight coefficients of the mixed molecular spectrum components on the basis of their reference samples are considered in this paper. It has been established that the presence of additive noise in the output mixed spectrum generates a noise component with an unidentified probability distribution law in the found weight coefficients. The power generated by the noise can be several orders of magnitude higher than the power of the output signal additive noise. It is shown that the use of numerical methods for suppressing this noise, which is not based on its statistical characteristics, in particular, the median filtration, expands the limits of SNR, in which the proposed method maintains efficiency.

Markovian Learning Methods in Decision-Making Systems

Abstract: The optimal decision-making task based on the Markovian learning methods is investigated. The stochastic and deterministic learning methods are described. The decision-making problem is formulated. The problem of Markovian learning of an agent making optimal decisions in a deterministic environment was solved on the example of finding the shortest path in the cell space. The mathematical formulation of the decision - making problem with deterministic and stochastic strategies based on recurrent estimation of criterion functions of utility of states and efficiency of options of actions of the agent was provided. The evaluation of criterion functions values takes place in real time on the basis of reinforced Q-learning and does not require a model of the environment, which is important for practical applications of decision-making in conditions of uncertainty. The algorithmic and software tools for the decision making modelling in uncertainty conditions are developed. The computer simulation results of decision-making process in cellular space are discussed and presented.

Forest Cover Type Classification Based on Environment Characteristics and Machine Learning Technology.

Abstract: Today, one of lot global problems is forest deforestation and monitoring. The article describes the system creation for forests control and monitoring. This work aims is to develop a model for the forest cover type determination based on environmental characteristics and machine learning as the currently developing project part “Monitoring the trees condition using drones”. The project aims is to simplify and partially automate the control and monitoring of trees using drones and machine learning to improve the forest situation. The task is to create a model for predict what trees types grow in the area based on environmental characteristics. Therefore, the main system will able to compare the existing values/characteristics with the predicted ones (which tree should normally there) to find discrepancies. Eventually, the main system will able to use this information to report and inform relevant staff and authorities. This work is based on a data set for learning system that includes observations of trees from four areas of Roosevelt National Forest in Colorado. All observations are cartographic variables (without remote sensing) from \(30\times 30\)-m forest areas. In total, there are more than half a million measurements. The work aim is the development of a forest cover types classification model depending on the environment and its characteristics.

Detecting Items with the Biggest Weight Based on Neural Network and Machine Learning Methods

Abstract: The article considers the development of a program for predicting areas on the design of the graphical interface, where users’ attention will be greatest. Current methods of improving graphic design in accordance with UX and software for getting UX information are considered. In this project the problem in complexity of creation of the graphic user interface for modern programs, web resources, and applications is considered. Namely, the problem is related to UX design, with errors that occur at the stage of product launch. Also considered methods to help avoid errors, software, using which, will help avoid difficulties and errors. The concept of developing a program for recognizing areas to which the user is most likely to pay attention in the first place is considered. This program aims to help designers to create a modern graphical user interface design that will guide the user to the right places and the user will spend a minimum of time and effort to interact and use the program. A method has developed to collect data on the behavior of the user’s eyes when using the program, during user interaction with the program interface. This method will create a database on the basis of which the program to predict the most visible areas will return the result.

The control agent with fuzzy logic

Abstract: In this paper a controller construction on the basis of the intelligent agent with fuzzy logic is considered. The algorithm of a fuzzy logic inference of the agent decision-making is described. Results of computer modelling of control process are received and analysed.

Sustainable Development by a Statistical Analysis of Country Rankings by the Population Happiness Level

Abstract: A statistical analysis of world rankings for the population happiness level was conducted to find ways to stimulate sustainable development. The research is based on data from an annual Gallup World survey called the World Happiness Report. Various data describing the population's happiness level through direct (GDP per Capita, Social support, Life expectancy) and indirect well-being indicators have been studied. The visualization methods, graphical mapping in Cartesian and polar coordinate systems and primary statistical processing of numerical data were used. We have used descriptive statistics, histograms, and cumulative constructions. Linear smoothing was performed, and the GDP per Capita rating trend was used to establish. Since the level of economic development and well-being of the population does not correspond to its happiness, it is recommended to include the implementation of the population's socio-historical, cultural and psychological traditions.

Matrix Stochastic Game with Q-learning for Multi-agent Systems

Abstract: A model of matrix stochastic game for decision making in conditions of uncertainty is developed. A Q-learning method is proposed for solving a stochastic game with a priori unknown payoff matrices. The formulation of the game problem is performed, the Markov recurrent method and the algorithm for its solution are described. The results of computer simulation of stochastic game with Q-learning are obtained and analyzed. In this paper, the ranges of changes in the parameters of the game Q-method are experimentally established to ensure the convergence to one of the Nash equilibrium points. As the value of the current win discounting parameter increases, the variance of current winnings reduces, and the order of change for the learning step decreases, and the convergence rate of the game Q-method increases.

Intelligent Analysis of Best-Selling Books Statistics on Amazon

Abstract: A data set of 550 elements was analyzed during the study, containing data on books: title, author, number of reviews, price, and rating. Smoothing methods were used to average local data for further forecasting, in which non-systematic elements replace each other. Clear graphs without sharp peaks were obtained. Based on which the methods of moving average, weighted moving average, exponential smoothing and median filtering, it was found that the books that became bestsellers in 2009-2011 did not have a high response rate. While in the following 2012 to 2019, there was an increase in the number of reviews; the average was 25,000 per book. Examining the correlation between Reviews and Prices, we found a relationship that indicates that books priced above $ 60 typically have no more than 10,000 reviews. Books priced from $ 40 to $ 60 have an average of no more than 20,000 reviews. Books up to $ 20 have different reviews, but mostly this value does not exceed 40,000, and those books that have more are the exception rather than the rule. The clustering method was applied to the 12 authors who wrote the bestsellers, and after normalizing the table and building proximity tables, we identified clusters.

Markovian Learning Methods in Decision-Making Systems

Abstract: The optimal decision-making task based on the Markovian learning methods is investigated. The stochastic and deterministic learning methods are described. The decision-making problem is formulated. The problem of Markovian learning of an agent making optimal decisions in a deterministic environment was solved on the example of finding the shortest path in the cell space. The mathematical formulation of the decision - making problem with deterministic and stochastic strategies based on recurrent estimation of criterion functions of utility of states and efficiency of options of actions of the agent was provided. The evaluation of criterion functions values takes place in real time on the basis of reinforced Q-learning and does not require a model of the environment, which is important for practical applications of decision-making in conditions of uncertainty. The algorithmic and software tools for the decision making modelling in uncertainty conditions are developed. The computer simulation results of decision-making process in cellular space are discussed and presented.