In the context of growing demands for rapid decision-making and in-depth analysis of complex dynamic systems – particularly when available data are limited and the involvement of experienced experts is either impractical or prohibitively expensive – the development of new methods for the construction of the model becomes especially relevant. The use of large language models (LLMs) as expert systems offers significant reductions in resource expenditure and accelerates the modelling of complex technical, environmental, and socio-economic systems. This study aimed to investigate and demonstrate the potential and capabilities of LLMs as expert systems in constructing cognitive maps. The article proposes and substantiates an architecture for the cooperation of LLM ensembles to formally generate vertices-variables and weight coefficients in cognitive maps, thereby enabling the automation of the modelling process without the involvement of human experts. A typical prompt for an LLM was decomposed into structural components: context description (D), model role instruction (R), instruction (I), conditions (C), and response format (F). A method for determining these components through expert-based analysis is proposed. A prompt system was developed to enable structured data processing and the identification of interrelationships among system elements. The practical effectiveness of the approach was demonstrated using a case study on forecasting water quality in the Sabarivske Reservoir near Vinnytsia. For most physicochemical indicators, the modelling showed low error rates (2.09-4.6%), even with a minimal amount of input data. The proposed method is promising for modelling and forecasting tasks in complex systems with limited data availability, particularly in environmental, socio-economic, and engineering contexts, where the speed of obtaining reliable results is critical for informed decision-making
No takes yet. Share an insight, caveat, or question.
В. Б. Мокін (2025) studied this question.