Search Papers | Poster Sessions | All Posters
Poster B10 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
A Generative Grammar for Automatically Designing Experiments on Human Learning and Decision Making
Maria Eckstein1, Kevin Miller1,3, Angela Radulescu2; 1Google DeepMind, 2Center for Computational Psychiatry, Icahn School of Medicine at Mount Sinai, 3University College London
The study of human learning and decision making has fascinated many researchers for a long time. As a result, the number of experimental paradigms in this field is large, and increasingly more sophisticated experiments are added continuously. While this multitude of approaches has provided innumerable insights, the resulting fragmentation has also led to wide-ranging contradiction in results, which have been difficult to resolve. We propose a method that leverages the strength of using multiple tasks to study a complex phenomena, while mitigating its disadvantages. Our method involves the specification of a family of tasks, defined by a procedural grammar that is based on a small number of task features. Rather than designing tasks individually, the grammar allows sampling them across the allowed space defined by the features. A dataset of human choices collected using this method is expected to reveal foundational insights about human learning and decision making that are generalizable and robust to task variations.
Keywords: reinforcement learning procedurally-generated tasks meta-learning