Abstract
This paper investigates a grid-representation-based approach to spatial cognition for intelligent agents, aiming to develop an effective neural network model that simulates the functions of the olfactory cortex and hippocampus for spatial cognition and navigation. Despite progress made by existing models in simulating biological nervous system functions, issues such as model simplification, lack of biological similarity, and practical application challenges remain. To address these issues, this paper proposes a neural network model that integrates grid representation, reinforcement learning, and encoding/decoding techniques. The model forms a grid representation by simulating the integration of grid cells in the medial entorhinal cortex (MEC) with perceptual information from the lateral entorhinal cortex (LEC), which encodes and retains spatial location information. By leveraging attractor networks, convolutional neural networks (CNNs), and multilayer perceptrons (MLPs), the model achieves the storage of spatial location and environmental information, as well as the construction of cognitive maps. The experimental results show that after using this model, the map generation accuracy increased by 15%, the navigation accuracy of the agent in complex environments by 20%, and the target localization error was reduced to less than 10%, demonstrating a significant overall performance improvement in the grid-based cognitive map construction.