States Representations with a Hierarchical Dependency in Reinforcement Learning

S. Kamdem, H. Iwasaki, H. Ohki, and N. Sueda (Japan)


Reinforcement Learning, Self Organizing Map, Au tonomous Agent.


In this research, we focus on the state space construction problem in cases where the state space exhibits some hi erarchical dependencies. In the case of the control of a robot’s motion, it is believed there exist a strong correlation between the current position and the current speed. For po sitions close to the goal area, it is necessary to define an area for slowing down whereas a high speed motion is con venient for distant positions. We conduct several experi ments to evaluate the effects of combining Self Organizing Map and grid-based representations to autonomously build such hierarchical structures. The results showed that with an extended acceleration range and a more detailed control, the hierarchically built state spaces permitted to achieve a smoother convergence, a better accuracy of learning and a more stable control.

Important Links:

Go Back