Affective and Human-Like Virtual Agents

Abstract

In Artificial Intelligence (AI) one of the technological goals is to build intelligent systems that not only perform human level tasks efficiently, but can also simulate and exhibit human-like behaviour. As the emphasis of systems is often placed on fulfilling functional requirements, AI systems are only intelligent at a machine level. Affective computing addresses this by developing AI that can recognize, understand and express emotion. In this work, we study the effects and humanness of emotionally cognizant AI agents within the context of the prisoner's dilemma. We leverage machine learning techniques and deep learning models in devising algorithms to map dimensional models of emotion to facial expressions for virtual human displays. Additionally, we utilize distributed representations for words to design a method for constructing affective utterances for a virtual agent in the prisoner's dilemma. We experimentally demonstrate that our methods for affective facial expression and utterance construction can be successfully used in AI applications with virtual humans. Thus, we design and build a prisoner's dilemma game application including the integration of a virtual human. We conduct two experiments to study and evaluate humanness of various agents in the prisoner's dilemma game. We demonstrate the effectiveness of our facial expression and utterance methods and show that an appraisal-based theoretic agent is perceived to be more human-like than baseline models.