Hryshko A. Hybrid machine learning methods in systems of intellectual processing of data.

The thesis is devoted to the development of hybrid machine learning techniques in intelligent systems for various applications (for example, trading systems and intelligent control systems) that improve decision-making strategies. The method for adapting the structure of technical indicators for the current state of the stock market when the trader combined RL-learning using genetic algorithms was propose. Hybrid method of inventory control with the use of stochastic dynamic programming and reinforcement learning technique that is compatible with non-separable criterion was propose. Modified method of neural network approximation of the Q-function RL-algorithm to allow for the correction of the configuration approximating MLP. The method for managing a dynamic object based on the replacement of states, using the predicted values for the signals reinforce previous suspension, the current value of the signal reinforcement was propose. The method is an extension of SARSA-algorithm and takes into account the evaluation of states far distant from each other. Improved structure prediction model, a learning algorithm which is based on application of neural network predictor filter, which is in contrast to the existing delivers high performance and quality predictions in unsteady and uncertain. The proposed model can be used to predict the trend signal reinforcement for Intelligent dynamic obektomi. The developed methods and software have been implemented and used for a number of practical implementations.