Performance of LQ-learning in POMDP Environments

抄録

In this paper, we propose a new type of LQ-learning to solve POMDP. In the POMDP environment, the agent cannot observe the environment directly. In the LQ-learning, in order to dicriminate partially observed states, the agent attaches label to each observation which perceived as the same ones. Unlike our previous LQ-learning, we make preparations of knowledge about the environment in advance. The knowledge is automatically acquired by Kohenen’s Self-Organizing Map (SOM), which provides the knowledge about state transitions to the agent. Then, LQ-learning agent attaches labels to observations with reference to a map obtained by SOM.