Optimization of Learning Cycles in Online Reinforcement Learning Systems

In reinforcement learning, since the learning cycles are set by the designer, this drawback greatly affects the learning efficiency. In this research, we adaptively change the learning cycles of online reinforcement learning systems to acquire a necessary and sufficient set of states for them. We used a growing self-organizing map to estimate the state for fast learning speed. In conventional methods, a calculation is performed more often than necessary, but in our proposed method, calculations that are unnecessary for learning are minimized based on the state transition. We demonstrate that the proposed method finishes learning quickly by the inverted pendulum task and, based on our experiment results, find that an approach that adaptively shortens or lengthens the learning cycle is suitable for reinforcement learning.