AUTOMATION OF SOLVING OPTIMIZATION PROBLEMS USING A HYBRID INTELLIGENT SYSTEM BASED ON LARGE LANGUAGE MODELS AND MIVAR NETWORKS

Kim R. I., Varlamov O. O., Afanasyev G. I.

This paper discusses the architecture of a hybrid intelligent system that combines the capabilities of large language models (LLMs) and mivar networks to automate the solution of discrete optimisation and scheduling problems. The proposed approach aims to overcome key limitations of existing language models, such as insufficient interpretability, inability to verify intermediate steps and lack of formal rigour of logical inference. The use of LLM allows us to interpret the natural-language description of the problem and automatically formalise it in the form of parameters, logical dependencies and production rules, while the mivar model provides a rigorous and reproducible logical inference. Special attention in the paper is paid to the study of the prospects of building more autonomous intelligent systems. In particular, the possibility of automatic formation of a mivar knowledge base without the involvement of an expert at the stage of model building is considered. The paper describes in detail the architecture of the proposed system, including modules of problem analysis, rule generation, verification, logical inference and solution formation. An experimental comparison of the performance of various large language models with the proposed hybrid architecture on a sample of typical optimisation problems is also carried out, which showed the advantage of the hybrid model in terms of accuracy and completeness of the solution.

Keywords:гибридные интеллектуальные системы, большие языковые модели, миварные технологии, миварные сети.