<p>In cloud computing, the dynamic task assignment problem (DTAP) is a task assignment problem (TAP) that involves simultaneous assignment of multiple tasks to multiple machines while the task set dynamically changes over time. To achieve a near-optimal solution in DTAP, the solver should be capable of considering not only the machine-specific or task-specific information, but also the relational information between tasks and machines. However, previous approaches based on deep reinforcement learning (DRL) for task assignment either did not consider relational information or considered it indirectly. Moreover, previous learning methods are not suitable to train a scheduling agent capable of making multiple assignments simultaneously. In this paper, we propose a novel architecture capable of comprehensively processing the relational information between tasks and machines, and a DRL framework to train a DTAP solver that is able to assign multiple tasks to multiple machines in a single timestep. We formulate the cloud computing scheduling DTAP as a problem of sequential decision-making based on the Markov Decision Process (MDP) and introduce a recurrent policy framework to solve the formulated MDP. We assessed our model using a real-world cloud computing scheduling dataset from Alibaba, comparing both makespan and energy consumption with other baseline methods. In a complicated scenario where tasks and machines possess varying specifications, our proposed method shows approximately 20% lower makespan and 13.2% lower energy consumption compared to state-of-the-art DRL methods.</p>