METHODS AND ALGORITHMS FOR TRAINING NEURAL NETWORKS BASED ON EXPERT KNOWLEDGE FOR EMERGENCY MANAGEMENT IN RAILWAY TRANSPORT

Abstract

In the context of increasing complexity and digitalization of decision-making processes, there is a growing demand for intelligent decision support systems capable of operating effectively in weakly formalized domains. Such domains are typically characterized by insufficient observational data, high uncertainty, and a significant role of expert knowledge. Under these conditions, traditional neural network training methods demonstrate limited applicability, which necessitates the development of alternative approaches focused on integrating expert experience into the learning process. The objective of this study is to develop a method for training artificial neural networks based on expert knowledge that enables the formation of a training dataset in the absence of sufficient statistical observations. To achieve this objective, existing approaches to neural network training were analyzed, and algorithms for generating training stimulus–response pairs, processing training data, and training neural networks using visual representations of situations were developed. The results of the study include the proposed expert-based neural network training method and a set of algorithms for data generation and preprocessing. It is demonstrated that a neural network trained using the proposed approach accumulates both formalized and implicit components of expert knowledge and can be applied in both “black-box” and “gray-box” modes, including knowledge extraction. In conclusion, the validity of the research hypothesis is confirmed, showing that effective neural network training in data-scarce environments is achievable through expert knowledge integration. The practical applicability of the proposed method in intelligent decision support systems is substantiated, and directions for further research are outlined.