Using an ethologically-inspired model to learn apparent temporal causality for planning in synthetic creatures

Inspired by recent work in ethology and animal training, we integrate representations for time and rate into a behaviorbased architecture for autonomous virtual creatures. The resulting computational model of affect and action selection allows creatures to discover and refine their understanding of apparent temporal causality relationships which may or may not involve self-action. The fundamental action selection choice that a creature must make in order to satisfy its internal needs is whether to explore, react or exploit. In this architecture, that choice is informed by an understanding of apparent temporal causality, the representation for which is integrated into the representation for action. The ability to accommodate changing ideas about causality allows the creature to exist in and adapt to a dynamic world. Not only is such a model suitable for computational systems, but its derivation from biological models suggests that it may also be useful for gaining a new perspective on learning in biological systems. The implementation of a complete character built using this architecture is able to reproduce a variety of conditioning phenomena, as well as learn in real-time using a training technique used with live animals