Regresar

Neural Agents with Continual Learning Capacities

Abstract:

The contemporary Artificial Neural Networks (ANNs) often suffer from catastrophic forgetting, where learned parameters are overwritten by new tasks. This paper presents a novel approach using a Reinforcement Learning (RL) agent with Continual Learning (CL) capabilities to navigate a visual robotic structure, achieving advanced proficiency in Tic-Tac-Toe. The system integrates a webcam for environmental perception, specialized neural blocks for feature extraction, and a communication bus linking self-taught agents with advisors. A knowledge protection mechanism prevents the loss of acquired parameters during new learning iterations. The methodology was validated on a physical robot, implemented with C++ and OpenCV, demonstrating its ability to retain knowledge and enhance gameplay, effectively emulating intelligent children’s learning strategies. The proposed system was tested in a real-world setting, achieving an average accuracy of 92% in task completion and demonstrating a 15% improvement in task retention over traditional methods.