强化学习引导的产品变更路径多目标差分进化算法

Reinforcement learning guided multi-objective differential evolutionary algorithm for product change paths

中文关键词  多目标优化  设计变更  差分进化  强化学习

英文关键词  multiobjective optimization  design change  differential evolution  reinforcement learning

中文摘要

      由于零部件之间复杂的关联关系, 产品设计变更效应的传播在所难免. 为降低产品设计变更带来的风险, 以变更对产品综合性能影响、变更经济成本和变更工期作为优化目标, 本文提出了一种强化学习引导的产品变更路径多目标差分进化算法. 首先, 建立问题的复杂产品网络模型, 揭示产品零件变更的传播机制; 接着, 引入变更传播强度, 间接评价零件变更对产品综合性能的影响; 同时, 考虑变更经济成本和变更工期指标, 建立产品设计变更传播路径的多目标优化模型; 进一步, 利用双深度Q-网络帮助种群在不同阶段选择适合的进化策略, 提出一种强化学 习引导的差分进化算法, 简称为DDQN-DE, 并通过上述算法求解最佳的产品设计变更传播路径. 最后, 以创维公司某型号电视机的设计变更问题为例, 并与已有算法进行对比, 实验验证了所提算法的有效性.

英文摘要

      Due to the complex interrelationships between components, the propagation of product design change effect is inevitable. To reduce the risk associated with product design changes, this paper proposes a multi-objective differential evolution algorithm guided by reinforcement learning to optimize the impact of changes on the product performance, economic cost, and change duration. Firstly, a complex product network model is established to reveal the propagation mechanism of component changes. Then, the change propagation intensity is introduced to indirectly evaluate the impact of component changes on the product performance. Meanwhile, considering the economic cost and change duration of design changes, a multi-objective model for optimizing design change propagation paths is established. Furthermore, utilizing dual deep Q-networks to assist populations in selecting appropriate evolutionary strategies at different stages, introducing a reinforcement learning-guided differential evolution algorithm, abbreviated as DDQN-DE, and using the above-mentioned algorithm to determine the optimal propagation path for product design changes. Finally, taking the design change problem of a TV product from skyworth company as an example and comparing with existing algorithms, the validity of the proposed algorithm is verified through experiments.