Growing up virtual : the computational lessons of development

Abstract

Cognitive development is one of nature's most important mechanisms for creating robustly adaptive intelligent creatures. From felids to oscines, developing animals are capable of learning in adverse environments with a reliability that often outpaces the current state-of-the-art in artificial intelligence (AI) The purpose of this thesis, therefore, is to examine how insights from cognitive development might be applied to the design of AI architectures. Starting with a targeted review of the ethological literature, I identify the key computational lessons of development, the fundamental conceptual insights that suggest intriguing new strategies for behavioral organization. These insights are then employed in the design of a developmental behavior architecture in which a hierarchical motivation-based behavior system is coupled to a distributed set of domain-specific learning tools. The architecture is deployed in a synthetic character (Hektor the mouse) whose challenge is to learn to play a competitive card matching game successfully against a human user. Evaluation of Hektor's performance on this task, at both qualitative and quantitative levels of description, reveal that the developmental architecture is capable of surmounting complex learning objectives in a novel and efficient manner. I conclude that the architecture presented here represents a valuable starting point for further consideration of developmental design principles.