Building Models of Human Behavior
In order to improve AI systems, I wish to imbue them with a greater understanding of human decision making and behavior using different forms of data. In the past, I have focused on modeling humans in video games and other virtual environments, although my work is not limited to these environments. In particular I have focused on using these models to create systems that generate dynamic interventions. These interventions include altering the AI of a computer controlled opponent and dynamically altering quests presented to the player in a 2-D role-playing game.
- Brent Harrison and David L. Roberts. Analytics-Driven Dynamic Game Adaption for Player Retention in a 2-Dimensional Adventure Game. InProceedings of Tenth Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE 2014). Raleigh, North Carolina. 2014. [pdf]
- Brent Harrison and David L. Roberts. Analytics-Driven Dynamic Game Adaption for Player Retention in Scrabble. In Proceedings of the Computational Intelligence and Games Conference (CIG 2013). Niagara Falls, Canada. Best Paper Nominee. [pdf]
Creating Believable Virtual Agents
Another major project I am engaged in involves using data to create believable virtual agents. There has been an increasing interest in integrating human feed back and demonstrations into machine learning algorithm in order to make them more accessible to non-experts. In this work, I am researching how stories, an unexplored source of human demonstrations, can be used to generate reward functions for reinforcement learning agents.
- Brent Harrison and Mark O. Riedl. Learning From Stories: Using Crowdsourced Narratives to Train Virtual Agents. Proceedings of the 2016 AAAI Conference on Artificial Intelligence for Interactive Digital Entertainment. Burlingame, California, 2016.
- Brent Harrison and Mark O. Riedl. Towards Learning From Stories: An Approach for Interactive Machine Learning. Proceedings of the AAAI Workshop on Symbiotic Cognitive Systems. Phoenix, Arizona, 2016.
- Mark O. Riedl and Brent Harrison. Using Stories to Teach Human Values to Artificial Agents. Proceedings of the 2nd International Workshop on AI, Ethics and Society. Phoenix, Arizona, 2016.
- Alexander Zook, Brent Harrison, and Mark Riedl. Monte-Carlo Tree Search for Simulation-Based Play Strategy Analysis. In Proceedings of the 2015 Foundations of Digital Games Conference. Pacific Grove, CA. 2015. Best Paper Nominee. [pdf]
- Using Stories to Teach Human Values to Artificial Agents – Georgia Tech Press Release. [link]
- Robots Read Books to Learn Right and Wrong – Newsweek. [link]
Improvisational Computational Storytelling
I am also exploring how deep learning techniques, specifically deep reinforcement learning, can be used for improvisational computational storytelling. Improvisational computational storytelling involves agents taking turns telling part of a story. In this work, I want to model this problem as a reinforcement learning problem in which an agent’s action space is the space of all possible natural language sentences that can be formed. Due to the sheer size of the state space, I am currently framing this as a deep reinforcement learning problem.
- Lara J. Martin, Brent Harrison, and Mark O. Riedl. Improvisational Computational Storytelling in Open Worlds. In Proceedings of the 2016 International Conference on Interactive Digital Entertainment. Los Angeles, California, 2016.
Explainable Artificial Intelligence
In many cases, the internals of a machine learning model or algorithm are a black box. Thus, it can be difficult for a human to determine the reasoning behind a model or agent’s behavior. In this work, I am seeking to train an agent that can ascribe human understandable meaning to its actions. To start with, I am examining how reinforcement learning policy information in games such as Frogger can be translated into human readable text describing why certain actions were taken in certain states.