Применение технологий генеративного искусственного интеллекта к задаче календарного планирования производства

Abstract

The paper addresses the problem of production scheduling under conditions of high dynamics, uncertainty, and resource heterogeneity. Traditional approaches based on centralized algorithms and ERP/MES systems often overlook individual characteristics of equipment and personnel, as well as poorly formalized technological constraints, which reduces the efficiency and adequacy of generated schedules. To overcome these limitations, an original hybrid architecture is proposed that integrates a multi-agent system (MAS) with generative artificial intelligence (GenAI) technologies. Building upon previously developed methodological foundations, a formal MAS model is defined as a tuple comprising resource agents, consumer agents, an ontological knowledge base, semantic constraints, and local/global optimization criteria. The key innovation lies in embedding a large language model (LLM) via the Retrieval-Augmented Generation (RAG) architecture, which enables the system to generate adaptive scheduling strategies, provide human-interpretable explanations, and simulate failure scenarios – while minimizing the risk of hallucinations. A two-stage agent interaction algorithm is proposed: in the first stage, a feasible schedule is constructed based on resource availability; in the second stage, the schedule is optimized through intelligent negotiations facilitated by generative AI. Experimental validation in a simulation environment modeling a machining shop floor demonstrates the effectiveness of the proposed approach. The results confirm that integrating generative AI into multi-agent systems creates a cognitive layer that enhances not only scheduling performance but also user trust – a critical factor for the successful deployment of AI in real-world production and management processes.