Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Investigating behavioural and affective cloning via imitation and reinforcement learning

Kapenekakis Antheas

Full record

---

Summary

In recent years, studying affect in computer-user interactions, video games, and even live streams has become increasingly popular. Objectively measuring the emotional experience of an audience has important implications in revenue generation and user retention. What has been relatively unstudied is user modeling by affect, especially in the context of video games. In this thesis, an initial framework and proof-of-concept for creating an affective agent is presented, which leverages user provided annotations to change agent behavior towards displaying a specific emotion, while trying to complete a behavioral objective.To create this agent, a tool for sourcing annotations was created and used to form a dataset with affective annotations. Then, the dataset was tested for validity by running supervised learning experiments. Using a form of Deep Q Learning, along with the dataset, a set of agents was created with each having a different objective. A primary Reinforcement Learning agent focused on completing the environment and each of the rest focused on maximizing an emotion. Lastly, the set of agents was combined in a ratio to form composite agents that focused on both affect and behavior, with certain combinations being successful at both.

Available Files

Services