Multi-objective Dynamic AGV Scheduling Method Based on Deep Reinforcement Learning

In industrial production, automated guided vehicles (AGVs) are widely used for material transfer and workpiece transportation to improve production efficiency. The growing demands of dynamic orders and switching production bring great challenges to the dynamic scheduling problem of AGV systems. To address the dynamic tasks with different optimization objective weight factors, a self-attention based multi-objective reinforcement learning (SAMORL) AGV dynamic scheduling method is proposed in this paper. At each rescheduling point, the self-attention based multi-objective deep dueling double Q-network (SAMOD3QN) is utilized to estimate the Q-values of task allocation actions on each objective respectively. Then in the action choosing process, the Q-values on different objectives are weighted according to the given weight factors. In this way, the task allocation policy achieves quick adjustment to different optimization objectives. Furthermore, the beam search algorithm is utilized to expand the search space of the optimal action trajectory according to the cumulative reward and estimated Q-value. The effectiveness and adaptability of the proposed dynamic scheduling method is illustrated by test examples based on stochastically inserted tasks.