Studying Behaviour Emergence in Multi-Agent Daily Life Simulation

Abstract

Digital worlds offer an excellent opportunity for dynamic storytelling. However, adding inhabitants to a digital world that can adjust to their environment and express interesting unscripted behavior is challenging. This thesis proposes a method to combine short-term planning with long-term optimization to spark emergent behavior and maximize agents’ fitness to the environment. We show that our approach can reliably demonstrate emergent behavior and, at the same time, improve agents’ survivability in their local environment. Moreover, we analyze the primary design factors affecting the planner and the optimizer. Lastly, we discuss the next steps we believe are effective in reaching authentic emergent narratives.