信息工程与人工智能学院知识库

作为现代服务业的核心组成部分，现代物流业依托数据驱动的物流配送网络，实现了商品从仓储到终端消费者的时效跃进。作为物流体系的核心环节，物流配送中心承担着货物分拣、存储、配送等一系列复杂任务。随着订单量的激增，如何提高物流配送中心的运营效率，成为业界关注的焦点。在这样的背景下，自动导引车（automated guided vehicle，AGV）以其高效、灵活的特点，逐渐成为物流配送中心的重要运输工具。然而，面对多负载AGV在复杂环境下的任务调度与路径规划问题的研究较少，且如何实现资源优化配置、降低运营成本，成为急需解决的问题。由此本文以物流配送中心为场景，以多负载AGV为研究对象，对AGV的任务调度和路径规划问题展开研究：

（1）本文在非支配排序遗传算法II（non-dominated sorting genetic algorithm-II，NSGA-II）算法上加以改进，得到MLNSGA-II算法以解决中小型物流配送中心场景下多负载AGV任务调度时其使用数量与配送效率之间存在边际效益的问题。以最小化AGV使用数量和最小化最大配送时间为优化目标的多目标调度模型，并考虑小车当前电量与负载行驶速度间的动态关系以及订单组数量等约束条件。进行仿真实验，得到MLNSGA-II算法在不同场景下求得最小AGV使用数量与最小-最大配送订单组总时间，以及订单任务和AGV之间的具体分配情况。仿真实验证明改进后的算法在解决该物流配送中心复杂任务调度问题时具有较好的可行性和易用性。对比实验结果显示，改进后的NSGA-II算法在求解时收敛速度更快，效果更优。

（2）本文提出了融入注意力机制的DDQN-A深度强化算法，以求解单AGV在静态环境中的路径规划问题。将配送中心抽象为栅格网络，AGV抽象为单智能体。在环境设置时，将智能体观察到的局部矩阵划分为三个通道：静态障碍物矩阵、历史访问矩阵、智能体所在位置矩阵，这些矩阵表示智能体的当前状态，将其输入到Q网络的卷积层中；为了平衡探索与利用之间的关系，在越界碰撞奖励、移动奖励、目标奖励的基础上，加入历史访问奖励。同时，将表示智能体与目标间关系的目标向量归一化后输入Q网络的全连接层，并结合Attention机制，动态调整智能体在不同时刻的环境下的观察重点。将融入注意力机制的Q网络与DDQN算法（double deep Q networks，DDQN）相结合得到DDQN-A算法。通过实验，将DDQN-A算法与Q-learning和双Q网络算法进行对比。结果表明，本文提出的DDQN-A算法在求解路径规划问题时具有更快的收敛速度和更好的稳定性，且路径更短更平滑。

关键词：任务调度 路径规划 多负载AGV 非支配排序 智能体

As a core component of modern service industry, modern logistics industry has achieved a significant leap in delivery efficiency from warehousing to end consumers by relying on data-driven logistics distribution networks. As a core link in the logistics system, logistics distribution centers undertake a series of complex tasks such as goods sorting, storage, and distribution. With the sharp increase in order volume, how to improve the operational efficiency of logistics distribution centers has become a focus of the industry. Against this backdrop, automated guided vehicles (AGVs) have gradually become an important transportation tool in logistics distribution centers due to their high efficiency and flexibility. However, there is a lack of research on the task scheduling and path planning of multi-load AGVs in complex environments, and how to achieve optimal resource allocation and reduce operational costs has become an urgent problem to be solved. Therefore, this paper takes the logistics distribution center as the scenario and multi-load AGVs as the research object to study the task scheduling and path planning problems of AGVs:

(1) This paper improves the non-dominated sorting genetic algorithm II (NSGA-II) to obtain the MLNSGA-II algorithm to solve the problem of marginal benefit between the number of AGVs used and the distribution efficiency in the task scheduling of multi-load AGVs in small and medium-sized logistics distribution centers. A multi-objective scheduling model is established with the minimization of AGV usage and the minimization of the maximum distribution time as the optimization objectives, and the dynamic relationship between the current battery level of the vehicle and the traveling speed under load, as well as the number of order groups, are considered as constraints. Simulation experiments are conducted to obtain the minimum AGV usage and the minimum-maximum distribution time of order groups, as well as the specific allocation of order tasks and AGVs under different scenarios. The simulation experiments prove that the improved algorithm has good feasibility and usability in solving the complex task scheduling problem of this logistics distribution center. The comparison experiments show that the improved NSGA-II algorithm converges faster and has better performance when solving the problem.

(2) This paper proposes the DDQN-A deep reinforcement learning algorithm integrated with the attention mechanism to solve the path planning problem of a single AGV in a static environment. The distribution center is abstracted as a grid network, and the AGV is abstracted as a single agent. In the environment setting, the local matrix observed by the agent is divided into three channels: the static obstacle matrix, the historical visit matrix, and the agent's position matrix, which represent the current state of the agent and are input into the convolutional layer of the Q network. To balance exploration and exploitation, historical visit rewards are added to the out-of-bound collision reward, movement reward, and target reward. Meanwhile, the target vector representing the relationship between the agent and the target is normalized and input into the fully connected layer of the Q network, and combined with the Attention mechanism to dynamically adjust the agent's observation focus in different environments at different times. The Q network integrated with the attention mechanism is combined with the DDQN algorithm (double deep Q networks, DDQN) to obtain the DDQN-A algorithm. Through experiments, the DDQN-A algorithm is compared with Q-learning and the double Q network algorithm. The results show that the DDQN-A algorithm proposed in this paper has a faster convergence speed and better stability when solving the path planning problem, and the path is shorter and smoother.

Keywords：Task scheduling；Path planning；Multi-load AGV；Non-dominated sort；Agent