Towards Improving Interaction with Virtual Agents

Virtual agents have become more and more common in modern life. We interact with computers agents on a day-to-day basis, whether it be personal assistant on a mobile phone or a chat representative online. In this thesis, we focus on improving the interaction of humans with virtual agents. We have designed an algorithm for computer strategic decision making that incorporate human decision making in adversarial settings. We extended Monte-Carlo Tree search to account for a prediction of human decisions. Furthermore, we have created an embodied virtual agent in collaborative settings. More specifically, we have design a interrogative training system with a virtual character that takes the role of the suspect. The virtual suspect's responses are derived from its psychological-based internal state which modulates depending on the interrogator's statements. Lastly, in order to improve the realism of the virtual suspect, we propose a method for synthesizing realistic hand motions while accounting for interactions. Realistic hand motion is particularly important for virtual agent that converse with their human counterparts. The method synthesize motion using a stack of convolutional auto-encoders. The method accounts for finger interactions with both static and deformable objects by solving a non-convex optimization problem with signed distance fields.

Game Theory,Monte-Carlo Tree Search,Embodied Agent,Hand Motion,Finger Motion,Character Animation,Deep Learning,Optimization,Signed Distance Fields,