Action Learning for Autonomous Virtual Actors

The goal of our work is to model believable virtual actors. Within the framework of learning by imitation, the virtual actor must be able to modify its prototypic behavior for miming in its imaginary world the observed behavior of a model which can be another virtual actor or an avatar controlled by a human operator. Each virtual actor has sensors for perceiving and effectors for acting and also a library of prototypic behaviors specified by FCMs for simulating it-self and others in its imaginary space. FCMs can give true perception & emotion. We propose a FCM-learning algorithm using meta-knowledge about learning in order to imitate a given behavior in real-time. The virtual actor autonomously selects training periods. This selection is inspired by neurophysiological experiments about active perception and hippocampus. The implementations undertaken in the multi-agent environment oRis are related to a sheepdog gathering sheep and to the adaptation of a virtual sailor to a given sailing ship. Keywords— Active Perception, Autonoy, Believable Agent, Fuzzy Cognitive Map, Imitation