- We tested humans, rats, and RL agents on a novel modular maze
- Humans and rats were remarkably similar in their choice of trajectories
- Both species were most similar to agents utilizing a SR
- Humans also displayed features of model-based planning in early trials
William de Cothi, Nils Nyberg, Eva-Maria Griesbauer, …,
E ́ le onore Duvelle, Caswell Barry, Hugo J. Spiers
De Cothi et al. use a novel open-field modular maze to test the spatial navigation abilities of humans and rats, comparing them to simulated reinforcement learning agents. They find that humans and rats are remarkably similar in their choice of trajectories, with both species displaying most similarity to agents utilizing a successor representation.
This paper is another example of the work being done that shows that rats and humans use similar methods the navigate by which are NOT based on Magnetism.
The paper refers to successor representation (SR) this is reinforcement learning and is the process by which an agent learns to predict long-term future reward.