DESIGN OF PICKER-TO-PARTS WAREHOUSE FULFILLMENT SECTIONS USING SURROGATE MACHINE LEARNING MODEL

Abstract

The design of picker-to-parts warehouse sections contains various decision parameters such as warehouse dimensions, routing policy, and storage assignment policy. Assessing the holistic importance of each decision parameter cannot be easily quantified due to their mutual interdependence. It is crucial to obtain this information and investigate the possible combinations of policies and warehouse specifications. To solve this problem, we use a surrogate machine learning model to simulate the warehouse conditions across varying pick list sizes. Seasonally varying demand and pick face requirements are also considered. A dataset derived from simulation is used to train various machine learning algorithms. The model uses the Monte Carlo method and average travel distance as the output parameter to evaluate performance. The Random Forest, Decision Tree, Gradient Boost, XG Boost, LightGMB, CatBoost, and a tuned Artificial Neural Network show the best performance in terms of error and fit. SHAP feature importance is calculated for interpretability analysis. Warehouse design practitioners and fourth-party logistic problems can easily adapt and deploy the developed warehouse simulation methodology and machine learning model to help with bid design in determining optimal warehouse parameters and policies.