ادغام الگوريتم‌هاي فراابتكاري و روش‌هاي يادگيري ماشين در بهينه‌سازي طراحي شبكه زنجيره تأمين

Integration of Metaheuristic Algorithms an‎d Machine Learning Methods in Supply Chain Network Design Optimization

مهندسي صنايع و آينده پژوهي

Supply chain network design is one of the most practical an‎d critical issues in industries. In a world where supply an‎d distribution decision-making constitute critical an‎d sensitive points in every industry, the optimization of the supply chain network has gained even greater significance. Various methods an‎d approaches have been employed for supply chain network optimization, an‎d specialists in this field continue to seek novel approaches for better efficiency. On one han‎d, exact solution methods perform very efficiently in terms of the accuracy of the final solution for supply chain network design problems. However, the computational complexity of problems sometimes rises to such an extent—due to the multiplicity of levels, various facilities, an‎d different constraints—that these methods are unable to provide a desirable solution within a reasonable time. Meanwhile, approximate solution methods can assist supply chain network designers, yet these methods also possess limitations regarding solution accuracy an‎d convergence across various iterations. In this research, an effort has been made to provide more efficient solutions for the supply chain network design problem in terms of objective function value, execution time, an‎d the number of iterations required to achieve convergence by integrating machine learning methods with metaheuristic algorithms. Accordingly, various forms of reinforcement learning methods, such as Q-Learning an‎d Deep Q-Learning, have been incorporated into the Genetic Algorithm, Mayfly Algorithm, an‎d Red Deer Algorithm. After executing each of the hybrid algorithms, the TOPSIS method was employed to compare the different algorithms. Numerical results indicate that the hybrid approach combining the Red Deer Algorithm an‎d Q-Learning reinforcement learning outperforms other methods in terms of objective function value an‎d solution convergence. This method can yield near-optimal solutions for supply chain networks in large-scale an‎d near real-world dimensions. With an increase in the number of facilities at each level of the supply chain, the proposed approach demonstrated the necessary robustness, showing complete improvement compared to the base Red Deer Algorithm an‎d an acceptable margin of error relative to the exact solution method. Furthermore, an examination of larger-scale problems revealed that increasing the problem dimensions does not increase the error of the proposed approach compared to the exact solution method, thereby confirming the efficiency of the proposed hybrid approach.

5

166933