Scholarlink

<p>Deep neural networks (DNNs) are a robust and versatile machine learning technique that offers a wide range of applications across various domains. They constitute effective and approximate models that can replace realistic models. However, designing, training, and refining the model poses a significant challenge to achieve the desired outcome. To address this challenge, metaheuristics are employed for DNN structure optimization to improve their accuracy and efficiency. The present work proposes a Harmony Search Optimization (HSO) technique to find the best DNN configuration to approximate a realistic FJSSP by considering makespan as a scheduling objective function. Different DNN hyper-parameters are adjusted according to the accuracy of the solutions, including the number of neurons in each hidden layer, activation functions for both hidden and output layers, and learning rate and momentum. A comparative study is investigated to analyze and evaluate DNN performances, considering three optimization objectives when evaluating validation data: minimizing the occurrence of low-precision solutions (<i>P</i><sub><i>S</i></sub>), minimizing Mean Absolute Percentage Error (MAPE) value, and minimizing the upper-bound value of Relative Errors (Max REs) between DNN-predicted makespan values and their corresponding real values in the validation set. The results obtained during the application of DNN to a case study representing FJSSP demonstrate the effectiveness, high accuracy, and fast responsiveness of the models. The three metrics evaluations provide DNN models with excellent computation time (more than 99% time saving compared to the simulation model). However, in terms of predicting DNN outcomes for all proposed production runs, the DNN model resulting from the Max REs evaluation presents a relative error of 2.22% or less, outpacing the other two assessments: <i>P</i><sub><i>S</i></sub> (≤ 3.04%) and MAPE (≤ 4.04%). In addition, it presents compact solutions, and the outliers do not exceed a relative error of 27.24%. This assessment offers a viable solution to attain more accurate DNN predictions, effectively addressing future approximation challenges.</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

Deep neural network parameter tuning using harmony search for a realistic flexible job shop scheduling

  • 期刊论文

  • 发表日期：

    2025年3月13日

查看全文
  • Bachir Mihoubi

摘要

Deep neural networks (DNNs) are a robust and versatile machine learning technique that offers a wide range of applications across various domains. They constitute effective and approximate models that can replace realistic models. However, designing, training, and refining the model poses a significant challenge to achieve the desired outcome. To address this challenge, metaheuristics are employed for DNN structure optimization to improve their accuracy and efficiency. The present work proposes a Harmony Search Optimization (HSO) technique to find the best DNN configuration to approximate a realistic FJSSP by considering makespan as a scheduling objective function. Different DNN hyper-parameters are adjusted according to the accuracy of the solutions, including the number of neurons in each hidden layer, activation functions for both hidden and output layers, and learning rate and momentum. A comparative study is investigated to analyze and evaluate DNN performances, considering three optimization objectives when evaluating validation data: minimizing the occurrence of low-precision solutions (PS), minimizing Mean Absolute Percentage Error (MAPE) value, and minimizing the upper-bound value of Relative Errors (Max REs) between DNN-predicted makespan values and their corresponding real values in the validation set. The results obtained during the application of DNN to a case study representing FJSSP demonstrate the effectiveness, high accuracy, and fast responsiveness of the models. The three metrics evaluations provide DNN models with excellent computation time (more than 99% time saving compared to the simulation model). However, in terms of predicting DNN outcomes for all proposed production runs, the DNN model resulting from the Max REs evaluation presents a relative error of 2.22% or less, outpacing the other two assessments: PS (≤ 3.04%) and MAPE (≤ 4.04%). In addition, it presents compact solutions, and the outliers do not exceed a relative error of 27.24%. This assessment offers a viable solution to attain more accurate DNN predictions, effectively addressing future approximation challenges.