Acquisition of global information from local observation with movement construction of internal state-representation under partial observability

Artificial agents operating in real world have to know the dynamics of interaction between the agent and the world. We focus on the role of mobility, which discovers the dynamics and brings global information to the agent with local observation ability. In this paper, we propose a method of behavior-acquisition by an agent operating in POMDP by autonomous acquisition of the internal state-representation while autonomously constructing observation space. In order to deal with partial observability of the environment, the agent constructs its internal state-representation that is a decision tree representing state-distinction according to its short-term memory of past observations and actions. The state-representation is constructed appropriately by successive segmentation based on real experience. We confirm that the proposed method, can acquire, navigation behavior of a real mobile robot by simulations and experiments.