Search Papers | Poster Sessions | All Posters
Poster C111 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Zero-shot spatial planning in humans and deep reinforcement learning agents
Denis C. L. Lan1 (), Laurence T. Hunt1, Christopher Summerfield1; 1University of Oxford
Humans are particularly good at planning ‘zero-shot’ (i.e. without prior experience of the environment), a skill that is especially apparent in spatial domains (e.g., navigating a new city). Zero-shot spatial planning likely depends on both ‘transition-based’ strategies that focus on connectivity between states and ‘vector-based’ strategies that focus on their relative spatial locations. We developed a novel behavioral paradigm to dissociate the use of the two strategies and show that human participants successfully arbitrate between them for zero-shot planning by using vector-based strategies to head in the general goal direction and transition-based strategies to fine-tune navigation near landmarks. Deep reinforcement learning models trained on the same task learn behavioral policies that are strikingly similar to that of humans. Analysis of the models’ learnt representations reveal the emergence of functional ‘modules’ that implement these strategies, each with distinct informational content, representational geometries, and activation patterns.
Keywords: spatial planning navigation cognitive maps