The Use of а Multilayer Perceptron for Predicting а Person’s Emotional State

The present work is devoted to the use of mathematical methods (machine learning model — a multilayer perceptron, and a genetic method of feature selection) to recognize a person’s emotional state based on his/her electroencephalogram data for further application of the results in creating software for brain-computer neural interfaces. The aim of the study is to train a multilayer perceptron based on EEG data for its further use in solving problems of recognizing a person’s emotional state. The dataset “Database for Emotion Recognition System — GAMEEMO”, containing recordings of electroencephalograms of 28 participants of the experiment, as well as data from questionnaires in which they noted their own feelings about the manifestation of certain emotions, their nature and intensity within the framework of two characteristics — valence and arousal, was taken as a basis. Participants played one of four computer games, each of them was supposed to provoke one of four emotional states: boredom, fear, calmness or joy, which are analyzed in this study. The features are the values of brain signals registered after a certain time interval during the passage of one of the four games by the participants of the experiment. The selection of features-electrodes is carried out by the genetic algorithm to increase the accuracy of prediction and identify the most important, from the point of view of the model, brain areas for decoding human states. As a basis for comparing emotional states (boredom, fear, joy, and calmness), the classification scheme of affective words proposed by J. Russell is used. As a result, the genetic method of feature selection made it possible to identify patterns in the location of the selected electrodes when recognizing emotions. The accuracy of the prediction was improved by analyzing precise frame ranges and identifying the time periods when the participants experienced this or that emotion, depending on the current game events.