<p>Recently, deep reinforcement learning (DRL) has been noted in Automated Guided Vehicle scheduling. Most studies intended the agent receiving state space as the input and the output is the regression values of mixed priority dispatching rules, being action space. Although this method is simple to operate, it limits the solution exploration and makes difficult to training, thus leading to bounded results. To address this point, the paper extends the solution space by defining each feasible task as an action directly. Simultaneously in order to accurately evaluate the selected action, we propose a novel state-action pair as an input to a Double Deep Q-Network. Six different scenarios are simulated in comparison with the previous method to verify the effectiveness of the proposed method. FlexSim-based simulation result is shown that the proposed method improve the makespan reduction rate by 3.72% more than the previous method. The proposed method can be applied to real-time optimization and intelligent decision making in the complex manufacturing system.</p>