Exploration of a Breakthrough Technology for Emergence of Symbol Processing by Neuro-based Reinforcement Learning Considering Time Axis

点击翻译按钮获取中文摘要

Efficient learning in a huge amount of spatio-temporal information holds the key to the emergence of higher functions in the real world. In this research, handling of time axis was especially focused on. A novel idea named "Causality traces" is propounded which judge the importance of events "subjectively" and are used for retrospective learning. Its learning performance exceeds that with the conventional method in value learning. Next, based on the idea that "concept" is formed from the difference of necessary motions, it is confirmed that discrete and abstract internal state representations are autonomously formed in a recurrent neural network through reinforcement learning. Furthermore, in autonomous communication learning, it was shown that information about target movement could be transmitted after learning. A novel perspective could be introduced to the handling of the time axis, but for the emergence of symbol processing, learning of dynamics should be further improved.