Case-based planning for large virtual agent societies

In this paper we discuss building large scale virtual reality reconstructions of historical heritage sites and populating it with crowds of virtual agents. Such agents are capable of performing complex actions, while respecting the cultural and historical accuracy of agent behaviour. In many commercial video games such agents either have very limited range of actions (resulting primitive behaviour) or are manually designed (resulting high development costs). In contrast, we follow the principles of automatic goal generation and automatic planning. Automatic goal generation in our approach is achieved through simulating agent needs and then producing a goal in response to those needs that require satisfaction. Automatic planning refers to techniques that are concerned with producing sequences of actions that can successfully change the state of an agent to the state where its goals are satisfied. Classical planning algorithms are computationally costly and it is difficult to achieve real-time performance for our problem domain with those. We explain how real-time performance can be achieved with Case-Based Planning, where agents build plan libraries and learn how to reuse and combine existing plans to archive their dynamically changing goals. We illustrate the novelty of our approach, its complexity and associated performance gains through a case-study focused on developing a virtual reality reconstruction of an ancient Mesopotamian settlement in 5000 B.C