A Computational Model of Flexible Maze Navigation Through Hippocampal Replay: Bridging Neuroscience and AI

Abstract

The hippocampus is widely recognized for its role in spatial memory and navigation, particularly through the phenomenon of neural replay. This study proposes a computational model that simulates hippocampal replay to support flexible navigation in dynamic maze environments. By integrating biologically inspired replay mechanisms with a reinforcement learning framework, the model was tested in three types of mazes—linear, Y-shaped, and open-field. The replay-based model significantly outperformed traditional models like DQN and A3C in success rate, path efficiency, and learning speed. The results underscore the importance of temporal sequence replay in forming goal-directed trajectories and adapting to changing environments. Comparisons with neuroscientific literature confirm the plausibility of the model, aligning with empirical findings on the predictive and retrospective roles of hippocampal replay in animal studies. This work offers a novel computational perspective on cognitive mapping and sets the foundation for developing more adaptive and human-like AI navigation systems.