Scholarlink

<p>Assembly sequence planning (ASP) determines the order in which the components of a product are assembled. Existing ASP methods have limitations in planning complex products with high computational costs and inefficient optimization. This research proposes a Q-Learning-based Genetic Algorithm (QLGA) to address these challenges. The GA is problem-dependent and typically requires tuning for specific problems. By incorporating Q-learning capabilities into GA, the performance of GA can be enhanced, particularly for solving NP-hard ASP problems. Different versions of GA and QLGA with various crossover operators and selection mechanisms are compared for evaluating the proposed approaches through case studies. The results show that the QLGA outperforms the traditional GA in terms of convergence rate and efficiency. ASP can be significantly improved when Q-learning techniques are combined with traditional GA methods, providing a versatile and efficient solution for product assembly planning. In case study 1, the QLGA achieved the shortest estimated assembly time (16.36s), which is 5.19% faster than the best GA baseline (ROX-GA, 17.26s), and 48.70–53.90% faster than TOX-GA (31.90s) and TCX-GA (35.49s) for a 21-part industrial drive product. In case study 2, the QLGA had the shortest assembly time (17.01s), 1.51% faster than the best GA baseline (ROX-GA, 17.27s), and 48.52–80.60% faster than TOX-GA (33.04s) and TCX-GA (87.69s) for a 19-component surface analyzer product.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reinforcement learning for self-adaptive genetic algorithm in assembly sequence planning

  • 期刊论文

  • 发表日期：

    2025年11月8日

查看全文
  • Nima Geran Malek,
  • Qingjin Peng

摘要

Assembly sequence planning (ASP) determines the order in which the components of a product are assembled. Existing ASP methods have limitations in planning complex products with high computational costs and inefficient optimization. This research proposes a Q-Learning-based Genetic Algorithm (QLGA) to address these challenges. The GA is problem-dependent and typically requires tuning for specific problems. By incorporating Q-learning capabilities into GA, the performance of GA can be enhanced, particularly for solving NP-hard ASP problems. Different versions of GA and QLGA with various crossover operators and selection mechanisms are compared for evaluating the proposed approaches through case studies. The results show that the QLGA outperforms the traditional GA in terms of convergence rate and efficiency. ASP can be significantly improved when Q-learning techniques are combined with traditional GA methods, providing a versatile and efficient solution for product assembly planning. In case study 1, the QLGA achieved the shortest estimated assembly time (16.36s), which is 5.19% faster than the best GA baseline (ROX-GA, 17.26s), and 48.70–53.90% faster than TOX-GA (31.90s) and TCX-GA (35.49s) for a 21-part industrial drive product. In case study 2, the QLGA had the shortest assembly time (17.01s), 1.51% faster than the best GA baseline (ROX-GA, 17.27s), and 48.52–80.60% faster than TOX-GA (33.04s) and TCX-GA (87.69s) for a 19-component surface analyzer product.