OPTIMIZING ORDERS GROUPING FOR PICKING IN A 3PL E-COMMERCE WAREHOUSE: A COMPARATIVE STUDY OF MACHINE LEARNING, LLM AND HEURISTIC APPROACHES

Background: This paper addresses a significant gap in logistics research by introducing and empirically evaluating a novel hybrid order batching method designed for 3PL e-commerce warehouses. The scientific goal was to compare the operational effectiveness of this method—integrating heuristic, machine learning, and a generative Large Language Model (LLM)—against ten other algorithmic approaches to identify a solution that best minimizes picking path lengths while ensuring robustness and fast processing times. Methods: The study was based on three months of operational data from a 3PL provider, covering orders processed via simultaneous multi-carton picking. The performance of eleven distinct methods was evaluated using key parameters, including average path length, its statistical distribution, and algorithm processing time. Results: The quasi-optimal heuristic method achieved the shortest average picking path, offering a strong trade-off between solution quality and runtime. However, the proposed hybrid approach proved to be the most flexible and robust against data variability. This research confirms the practical feasibility of using an LLM for real-time batching decisions, a relatively unexplored application in logistics. Conclusions: The study concludes that the adaptive, multi-stage structure of the proposed hybrid method significantly enhances efficiency in dynamic warehouse environments. This research represents one of the first attempts to empirically validate the application of a generative language model within a hybrid batching methodology for 3PL e-commerce logistics, confirming its practical potential and contribution to adaptive logistics management.