Abstract
In recent years, a multi-agent robot system (MARS) utilizing reinforcement learning and transfer learning has been deployed in real-world situations. Autonomous agents of MARS obtain behavior autonomously by multi-agent reinforcement learning method, and a transfer learning method enables to reuse the knowledge of other robot's behavior such as cooperative behavior. Those methodologies, however, have not been fully discussed systematically. Hence, we propose the knowledge co-creation framework leveraging the transfer learning and a cloud computing. Until now, we developed a Hierarchical Transfer Learning (HTL) as core technology of a knowledge co-creation and indicated effectiveness of the HTL in a dynamic multi agent environment. However, an effectiveness of our proposed HTL depends on transfer rate and approximation accuracy of knowledge generated by artificial neural network. In this paper, we evaluate the effectiveness of the HTL in new artificial neural network settings.
