Selforganizing memory: active learning of landmarks used for navigation.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Robotics

Background:

• Autonomous agents require sophisticated memory systems for navigation and homing.
• Existing memory architectures may not adequately support self-organized learning of spatial relationships.

Purpose of the Study:

• To propose and evaluate a novel memory architecture for agent homing behavior.
• To demonstrate autonomous learning and utilization of spatial information using landmarks.

Main Methods:

• A self-organizing neuronal memory architecture comprising a pre-processor, distributor net, and recurrent net.
• Integration of Hebbian learning, delta-rule, and active explorative learning.
• Testing stability through landmark removal/shifting and introduction of sensory/weight noise.

Main Results:

• The proposed architecture enables agents to autonomously learn homing behaviors.
• The system demonstrated stability and robustness against landmark perturbations and noise.
• Performance showed comparability to biological experimental findings.

Conclusions:

• The developed memory architecture effectively supports homing behavior in artificial agents.
• The self-organizing and explorative learning mechanisms are key to its success.
• The architecture offers potential insights into biological memory systems, such as insect mushroom bodies.