Task-oriented reinforcement learning for continuous tasks in dynamic environment

This paper presents a more realistic way of learning for non-episodic tasks of mobile agents, in which the generalized state spaces as well as learning process do not depend on the environment structures. This work has two main contributions. First, the proposed task-oriented reinforcement learning allows the agent to use several Q-tables based on the type of subtasks that greatly reduces the dimensionality in state spaces. Second, the use of relative information of the environment topology makes the system capable of working in dynamic, environment continuously.