Abstract
Recently there has been an increased attention to the episodic memory research in the community studying intelligent virtual agents. An episodic memory can exploit a hierarchical representation, in which higher layers present increasingly more abstract tasks.
During storage, an agent with episodic memory can directly access these abstractions from its own behavioral representations or from behavioural representations of agents being observed, provided the representations in question are hierarchical. However, the abstractions are lacking when observing behaviour of human users.
In this paper we propose a way in which our already existing computational model of episodic memory can be extended by an ability to incorporate knowledge about abstract tasks of human users. We have tested ID3 classification algorithm and hidden Markov model for reconstructing these tasks from observations.
Our approach was tested in a simple scenario with one agent situated in a 3D environment of the game Unreal Tournament.