Generativní agenti pro simulaci sociálního chování

König, Matúš

Simulating believable human behavior has been a long-standing chal- lenge in artificial intelligence. Large language models (LLMs) have shown emergent capabilities in approximating human behavior, making LLM-based generative agents an interesting approach to social simulation. Although previous research has shown promising results, issues limiting widespread adoption persist. This thesis addresses these shortcomings by introducing a highly customizable and scalable framework. We introduce several fea- tures, including memory management, BDI planning, the selective forgetting mechanism and cost-saving measures, creating a practical toolkit for social research. We validate the framework on a synthetic social network and a real-world dataset from the Love Island reality show. Through a series of experiments, we demonstrate social and rational agent behavior. We find the best configuration of framework parameters and use statistical analysis to show improved results. The experiments confirm that the framework suc- cessfully simulates information spread and that agents behave rationally in game-theoretic scenarios. Experiments on the Love Island dataset suggest that romantic partner prediction remains an open challenge, showing our ap- proach's limits. Together, these results demonstrate that our framework is a...

guest :: login Digital Repository
		Search		Submit		Help		About