MULTI-AGENT SYSTEMS FOR SIMULATION OF COMPLEX SOCIAL INTERACTIONS NPC

Abstract

Multi-agent systems for simulating complex social interactions of NPCs. The article investigates the architectural foundations of multi-agent systems designed to model complex social interactions among non-player characters (NPCs) in open-world video games. Key components of contemporary generative agent architectures are examined in detail, including observation streams that capture environmental events in natural language, hierarchical memory systems with reflection mechanisms that synthesise episodic experiences into higher-order generalisations, and adaptive planning modules that generate and dynamically revise daily action plans in response to changing contexts and social stimuli. The study draws upon landmark research such as the Generative Agents project (Stanford University, 2023), the Voyager open-ended embodied agent, and the OASIS large-scale social simulation platform to substantiate the transition from scripted NPC behaviour models – prevalent in commercially successful titles including The Elder Scrolls V: Skyrim, The Witcher 3: Wild Hunt, Mass Effect, and The Sims – toward autonomous agent architectures in which each character possesses individual goals, affective states, persistent memory, and the capacity for long-term strategic planning. A conceptual model is proposed for converting individual NPC plans into structured quest chains for the player, grounded in a five-stage cycle: plan generation, conflict analysis, conflict classification by type (resource-based, territorial, social, ideological), quest structure generation through the integration of Hierarchical Task Networks with LLM-driven narrative framing, and feedback incorporation whereby quest outcomes update the world state and trigger new agent plans. The model identifies points of intersection between agent intentions as natural sources of emergent quests, thus shifting the origin of game narrative from authorial scripting to organic inter-agent dynamics. Special attention is devoted to the scalability of such systems, computational cost analysis, the problem of maintaining narrative coherence when dozens of autonomous agents operate simultaneously, and hybrid architectures that combine LLM-based strategic planning with behaviour trees for tactical execution. The conclusions demonstrate that integrating multi-agent planning with procedural quest generation opens realistic prospects for creating game worlds with genuinely organic and unpredictable event development, while acknowledging that computational costs, hallucination risks, and coherence maintenance remain key challenges requiring further research.