| Summary: | This thesis puts forward a computational framework that can be used by embodied artificial agents (and in particular autonomous robots) for ontogenetic development. The research investigates methods, endowed with which, an embodied agent can develop control structures for increasingly complex and better adapted behaviour, explicitly and incrementally from its history of interaction with its environment. The temporal horizon of an agent is extended so that past experience can be self-organized into a developing structure that can be used to anticipate the future and act appropriately in environments where state information is incomplete, such as a social environment. A formal definition of sensorimotor experience is given, and Crutchfield’s information metric is used as the basis for comparison of experiences. Information metrics are demonstrated to be able to characterize and identify time-extended behaviour. A definition of a metric space of experiences is followed by the introduction of an architecture that combines this with environmental reinforcement as the basis for a system for robot ontogeny. The architecture is demonstrated and tested in various robotic and simulation experiments. This thesis also introduces the early communication game “Peekaboo” as a tool for the study of human-robot interaction and development. The interaction history architecture is then used by two different robots to develop the capability to engage in the peekaboo game.
|
|---|
