PhD student position at the School of Electrical Engineering and Computer Science at KTH Royal Institute of Technology.
This project addresses the question of how to elicit and incorporate human feedback in the development of autonomous agents, with a specific focus on the design of believable Non-player Characters (NPCs) in video games. While NPCs typically follow pre-scripted actions or dialogue that can affect gameplay, the gaming industry has a large interest in using AI to increase the autonomy and believability of these characters. We aim to make the following contributions:
- Develop novel algorithms that can align human feedback with existing models in an efficient manner. This includes (a) extending existing active and preference learning methods for determining which queries to elicit human input to maximize learning, as well as (b) developing novel human-machine interaction modalities to elicit feedback and obtain as much information as possible from each query.
- Establishment of appropriate benchmarks for evaluating subjective aspects of the learned behaviors (e.g., believability, congruency with the character type/context, etc.), beyond the typical metrics used in task-oriented learning.
- Demonstrate and evaluate our approach with real human feedback in scenarios that can highlight the significance of the developed methods beyond gaming applications.
The project is financed by the Swedish AI-program WASP (Wallenberg AI, Autonomous Systems and Software Program) which offers a graduate school with research visits, partner universities, and visiting lecturers. Although this is an academic research position, the research will be conducted in close collaboration with Electronic Arts.