Research of Crowd Behavior Simulation Methods Based on Relationships and Emotional Evolution

Simulating believable multi-agent behavior remains a significant challenge, particularly in endowing agents with intrinsic motivations and the capacity for complex, emergent social interactions. This paper presents an integrated framework that addresses these challenges by synergizing generative agents, powered by Large Language Models (LLMs), with a rich semantic environment and sophisticated internal state models. The framework leverages scene semantics derived from Building Information Modeling (BIM) to facilitate meaningful agent-environment interactions, employs the OCEAN personality model to instill agent heterogeneity, and uses a dynamic emotion model to drive internal state evolution. A novel emotion-based neural A* algorithm is introduced to translate these internal states into plausible, non-optimal pathfinding decisions. Experimental results, including quantitative analysis and ablation studies, demonstrate the framework's capability to generate agents that exhibit (1) human-like path choices influenced by their dynamic emotional states, and (2) emergent social behaviors, such as forming unscripted relationships, driven by a persistent memory and reflection mechanism. This work contributes a significant step towards creating more autonomous and lifelike virtual agents for complex simulations.