An Overview of the Action Space for Deep Reinforcement Learning

摘要

In recent years, deep reinforcement learning has been applied to tasks in the real world gradually. Especially in the field of control, reinforcement learning has shown unprecedented popularity, such as robot control, autonomous driving, and so on. Different algorithms may be suitable for different problems, so we investigate and analyze the existing advanced deep reinforcement learning algorithms from the perspective of action space. At the same time, we analyze the differences and connections between discrete action space, continuous action space and discrete-continuous hybrid action space, and elaborate various reinforcement learning algorithms suitable for different action spaces. Applying reinforcement learning to the control problem in the real world still presents huge challenges. Finally, we summarize these challenges and discuss how reinforcement learning can be appropriately applied to satellite attitude control tasks.