Control With Prametrised Actions

In reinforcement learning research, most papers focus on the behavior of agents in discrete or continuous action spaces. However, when training agents to play video games, it is often necessary to deal with composite actions that have both discrete and continuous components. This type of task is referred to as "Control with Parameterised Actions," which aims to design algorithms that enable agents to handle discrete decision-making and continuous parameter optimization simultaneously, thereby achieving efficient learning and execution in complex environments. The application value of this task lies in enhancing the adaptability and flexibility of agents in multi-modal interactive environments such as games.