Extraction of neural network controller rules in the construction of cognitive automatic control systems

Abstract

Cognitive technologies are included in one of the most «intelligent» sections of the theory of artificial intelligence. A special place in intelligent systems is occupied by precedent-based learning based on the identification of general patterns from particular empirical data, which is mainly implemented in artificial neural networks (ANNs). Due to their structural features, ANNs are successfully used for the synthesis of nonlinear controllers in automatic control systems. However, the neural network algorithm for generating results, which is implicit for the user, gives rise to the problem of trusting its conclusions when solving real practical problems. In this regard, the problem of establishing the «transparency» of the internal ANN algorithm is relevant. The purpose of the research is to increase the reliability of the functioning of neural network controllers in the construction of cognitive automatic control systems. The paper analyzes the existing approaches and methods for the interpretability of ANN results, considers the well-known methods of formalizing neural network algorithms that make it possible to describe the rules for ANN functioning. A method for extracting the rules of a neural network controller based on granular computations is proposed. As information granules, sets of input features similar in their properties, combined into clusters, are taken, which is achieved by integrating self-organizing layers into layers of multilayer ANNs. The applicability of the proposed solutions is shown on the example of the synthesis of an automatic control system with non-linear characteristics. A neurocontroller has been constructed that surpasses the known solutions of this problem in its characteristics, and also has the ability to verbally represent the rules of its operation. The practical significance of the proposed solutions is the construction of a causal relationship between the input data sets and the generated output signal of the neurocontroller, their representation in the form of a set of rules. This provides interpretation of the neural network algorithm in terms of building a cognitive automatic control system. The proposed method for extracting the rules of a neural network controller can be used in the methods of analysis and synthesis of intelligent automatic control systems, about which appropriate recommendations and suggestions are given.