When robots weep : a computational approach to affective learning

Abstract

This thesis presents a unified computational framework for the study of emotion that integrates several concepts and mechanisms which have been traditionally deemed to be integral components of intelligent behavior. We introduce the notion of affect programs as the primary theoretical constructs for investigating the function and the mechanisms of emotion, and instantiate these in a variety of embodied agents, including physical and simulated robots. Each of these affect programs establishes a functionally distinct mode of operation for the robots, that is activated when specific environmental contingencies are appraised. These modes involve the coordinated adjustment and entrainment of several different systems-including those governing perception, attention, motivation regulation, action selection, learning, and motor control-as part of the implementation of specialized solutions that take advantage of the regularities found in highly recurrent and prototypical environmental contingencies. We demonstrate this framework through multiple experimental scenarios that explore important features of the affect program abstraction and its function, including the demonstration of affective behavior, evaluative conditioning, incentive salience, and affective learning.