Human-like Task Adaptation in Artificial Agents with `What' and `Where' Representations
Arthur Juliani, Margaret E. Sereno · 2023
Humans are able to skillfully navigate their environments even when aspects of the environment unexpectedly change over time. Using a realistic virtual environment, we study the ability of both human participants and artificial agents to adapt to changes in a navigation task over time. We find that humans demonstrate a hybrid decision-making strategy, whereby they are able to better adapt to changes in superficial statistics of the environment or to goal location than they are to environment structure. To model this behavior, we compared a set of artificial agents utilizing state-space information from a dual-stream world model containing distinct learned `what' and `where' representations. We find that agents utilizing a `what' representation are able to most quickly learn the task, while agents utilizing a `where' representation as the basis of their learned policy are most consistent with the human behavior when task changes take place. This highlights the role that representation learning plays in downstream behavior, even when the underlying behavioral algorithm remains fixed.