I'm pleased to share our (Najjartabar Bisheh, Nasiri, Esmaeili, Davoudpour, and Chang) 2022 published article in the International Journal of System Assurance Engineering and Management entitled: A new supply chain distribution network design for two classes of customers using transfer recurrent neural network. Those colleagues who are interested in reading and downloading this paper, can use the following link or contact me:
https://link.springer.com/article/10.1007/s13198-022-01670-w
Abstract: Supply chain management integrates planning and controlling of materials, information, and finances in a process which begins from suppliers and ends with customers. Optimal planning decisions made in such a distribution network usually include transportation, facilities location, and inventory. This study presents a new approach for considering customers’ differentiation in an integrated location-allocation and inventory control model using transfer recurrent neural network (RNN). In this study, a location and allocation problem is integrated with inventory control decisions considering two classes of strategic and non-strategic customers. For the first time, a novel transfer RNN is applied to estimate parameters in order to reach to a near optimal solution. The proposed mathematical model is multi-product, single-period, multi-transportation mode, and with multilevel capacity warehouses with two classes of customers based on a critical level policy. The transfer RNN approach is used to transfer knowledge from a similar domain to the problem domain in this study. The performance result is compared with the condition when no transfer learning approach is applied. The exact calculation method is demonstrated for small scale instances while hybrid meta-heuristic algorithms (Genetic and Simulated Annealing) developed for real size samples. Finally, a sensitivity analysis is carried out for different instances to evaluate the effect of different indexes on the running time and total cost value of the objective function.
Keywords Supply network design Customer classification Recurrent neural network Transfer learning Inventory management Mathematical programming
In our distribution network, different customers (in terms of purchase volume and corporate strategies) may not carry the same significance to a producer and each customer may require a unique supply policy due to demand uncertainty. This study examines impacts of customer demand fulfillment in SC design components with respect to different importance of customers, demand uncertainty, transportation variation, limited capacity of warehouses, and backorders. For the first time, a transfer learning technique in neural networks is deployed to transfer information from a similar domain and predict important parameters such as demand and improve mathematical modeling performance by having better parameters estimation. The mathematical model is proposed and then two meta-heuristic algorithms of classic genetic algorithm (GA) and hybrid of genetic and simulated annealing (SA) is employed to solve a general model. According to the nature of meta-heuristic algorithms, a verification is done via analysis of variance (ANOVA) to ensure that the results are statistically significant.
Conclusion: In this study a distribution network model for a three echelon capacitated SC with uncertain customer demands and special respect to customer differentiation was developed. This paper with a critical review to previous studies on customer differentiation, introduced a new approach for modeling differentiation of customers into two classes of strategic and non-strategic ones by using a critical level of (C, S). Most of the studies are making an assumption that important parameters such as demand could be represented with a fixed distribution function. However, we know demand in each period might be different and needs to be estimated, therefore, this study used a novel transfer RNN approach to predict and estimate demand and place it in the mathematical model. The presented model was a mixed integer non-linear model which is placed in category of NP-hard problems and could not be solved with an exact method easily for real size problems with a reasonable running time. Therefore, the model was solved with two meta-heuristic algorithms. First, it was solved with a classic GA and then a hybrid of GA-SA is used where SA used for local search. For evaluating efficiency of algorithms, ANOVA was used which showed in small-size instances, with p-value & 0.1, the differences among both algorithms are not significant. The p-value for large-size instances was 0.5 which showed again there was no significant difference among algorithms. Finally, sensitivity analysis showed number of products and level for warehouses has the most important effect on running time and increase level for potential warehouses cause a great decrease in total costs. This study can be expanded to several directions such as considering N classes of customers or demand dependency. Another potential area of research is considering the problem as a multi-objective system to reduce delivery time and improved customer service level.