Hierarchical information-extreme machine learning of hand prosthesis control system based on decursive data structure

Abstract

The article considers the machine learning method for a hand prosthesis control system that recognizes electromyographic signals with a non-invasive recording system. The method was developed within the informationextreme intelligent data analysis technology framework to maximize the system’s information capacity during machine learning. The method is based on adapting the input information description to maximize the probability of correct classification decisions, similar to artificial neural networks. However, unlike neural-like structures, the proposed method was developed within a functional approach to modeling cognitive processes of natural intelligence formation and decision-making. This approach allowed the recognition system to adapt to arbitrary initial conditions of electromyogram formation and flexibility when retraining the system by expanding the alphabet of recognition classes. The decision rules formed by the results of information-extreme machine learning were characterized by high efficiency as an essential indicator of an intelligent prosthesis. The distinctiveness of the developed method from known machine learning methods was in applying a hierarchical data structure as a decursive binary tree, which allowed for transitioning from multi-class machine learning to two-class learning for each stratum of the decursive tree. The modified Kullback–Leibler information measure was the optimization criterion for machine learning parameters. The proposed hierarchical information-extreme machine learning method was implemented using electromyographic biosignals of cognitive commands for six finger and hand movements as an example.