Multi-agent reinforcement learning
A juvenile zebra finch tries to match its inventory of vocal gestures to that of an adult tutor. We hypothesize that the juvenile computes a vocabulary reward R. Our goal is to estimate this function R from observations.
Reinforcement learning has been a successful framework for understanding how animals and humans choose actions to maximize future reward. However, some actions such as vocal gestures used in speech and animal communication are best seen not as immediate responses to the environment, but as constituents of an action vocabulary that provides flexibility of expression thanks to combinatorial arrangement of action sequences. Little is known about how an action vocabulary can be learnt from a simple underlying scalar reward.
Project
To gain insights into reinforcement learning of action vocabularies, we trained young zebra finches to re-learn a song vocabulary. We fit developmental song trajectories with a multi-agent reinforcement learning (MARL) model comprising one agent for each vocalization type (syllables and calls).
The interest of this modeling effort is to relate complex natural behaviors to neural mechanisms of song learning instantiated by dopaminergic neurons and in the potential applicability of our findings to human language.
Collaborators
This project is a collaboration between our lab and that of Prof. Dina Lipkind at CUNY and Dr. Hazem Toutonji at the University of Nottingham.
Student Project
If you are interested in this project for an MSc Thesis or Semester Project, please get in touch with Anja Zai (zaia (at) ethz.ch).