Abstract
This paper proposes a method for acquisition of behavior capable of multiple tasks by an agent with non-predesigned and continuous observation space in partially observable environment. In the proposed method, the agent discriminates the ongoing task from other tasks using experience acquired in acquisition process of behaviors corresponding to the tasks, and applies knowledge for the corresponding task. In order to cope with lack of clues for task-identification and bad influence on the task-execution behavior by the task-identifying behavior, additional learning process is executed as the need arises. Simulation results shows the proposed method can build a suitable behavior for multiple tasks in a simple partially observable grid-environment.
