Abstract
Intelligent Virtual Agents (IVAs) are interactive characters controlled by software agents that exhibit human-like qualities. When interacting with human participants in virtual communities, an IVA is expected to be intelligent, as well as believable. A major factor that determines the intelligence and believability of an IVA is its ability to be aware of its physical and social environments.
An important aspect of physical environment awareness is the ability of an agent to refer at an abstract level to the interactions of entities (avatars and objects) in its surrounding environment. However, sensor data received from many virtual worlds describe the agent’s environment at the level of objects and avatars. This has introduced an information representation gap between virtual worlds and IVAs implemented using agent systems such as those based on the BDI architecture.
In this thesis, the dynamic virtual environment of an IVA is described using entity relationships that abstractly describe the state of the environment, communication actions included in the messages exchanged in chat channels, complex events that take place, and situations that are unfolding. An abstract model of virtual environments is formulated to describe these different virtual environment elements, and their interrelationships. A set of data inference mechanisms is identified, which interprets the low-level sensor data received from virtual worlds into the aforementioned elements of dynamic virtual environments. A framework implementing these data inference mechanisms is used to process sensor data received by Jason agents deployed in the popular multi-purpose virtual world Second Life.
With respect to improving social environment awareness of an IVA, an important aspect that has not been considered in previous research is how an IVA can be aware of the fulfilments and violations of its social expectations. A hindrance to enable an IVA to monitor for its expectations is the lack of such mechanisms in common agent frameworks such as those based on the BDI architecture. As a solution to this problem, this thesis integrates the concept of expectation monitoring with the BDI agent paradigm. An interface is defined that can be used by BDI agents to delegate monitoring for their expectations to monitoring processes outside the agent system. This interface is currently implemented for the Jason BDI agent development platform.