An adaptive optimization framework for manual order picking

Order picking (OP) is a key process in warehouse management: it is defined by a set of tasks that are performed to retrieve products from inventory with the goal of fulfilling customer orders in an accurate and timely manner. OP efficiency can be pursued on different levels. From a strategic perspec-tive, long-term shifts towards new and, most likely, more promising direc-tions and/or programs could be made by introducing new infrastructure and/or technology, which normally comes with a great cost. On the other hand, directions or programs could be translated into specific tactical actions, for instance, by matching existing infrastructure and technology with ade-quate solutions in terms of layout, organization and resource assignment and deployment. As for the operational level, company policies and practices could be conceived and/or tuned to ensure that daily activities are carried out efficiently with respect to, for instance, item storage/retrieval and handling. Amongst all the different options available, manual order picking in a rack-based facility is one of today’s most popular and versatile warehouse pro-cesses: it is also the most costly and labor-intensive in logistics organization. In light of these specific traits, optimizing manual OP compels to improve both planning and operational efficiency, while also accounting for human presence. Under the assumption that strategic investments are to be ruled out for eco-nomic reasons, we present an adaptive optimization framework based on a persons-to-goods principle in which OP efficiency including human well-being is considered on both the tactical and operational levels. Specifically, the proposed approach allows: i) assigning picking locations to items accord-ing to the “best” weight or physical size of the goods in order to minimize human fatigue and, thus, related errors and/or accidents [tactical level]; ii) as-signing picking lists to (free) order pickers according to utilization-guided rules such as the most idle or the longest idle worker in the current shift [op-erational level]; iii) setting tolerance intended as a time limit beyond which assigning picking lists to workers before shift change or at the end of the day is not possible [operational level]; iv) fixing worker behavior when the target picking position of a list is empty and the assigned item requires replenish-ment from the corresponding storage position, as in “stop-and-queue” or “skip-and-go” policies [operational level]. The above tactical and operational features are both represented in the pro-posed solution framework. In particular, a mathematical programming (MP) model for the location assignment (tactical) problem is embodied in an opti-mization component, while a discrete-event simulation component mimics (operational) actions which are triggered by and, in turn, trigger specific events. The backbone for generating and exploring neighboring solutions ac-cording to the objective function defined in the optimization model is an it-erative local search (ILS) procedure. At every iteration of the search proce-dure, simulation is used to evaluate the best solutions with respect to the or-der picking process. Therefore, simulation guides the search procedure. Based on the fixed computational budget, the framework adapts to consider either a sample (S) or fully enumerated (FE) set of solutions for small, medi-um and large-sized instances. Both variants return solutions that, in quality and on average, are within 1% of the optimal solution returned by solving the MP formulation in CPLEX 20.1.0. As for quantity, the performance tests focus on the (average) solution time required by the two variants of the ILS-based heuristic to solve each group of instances. The difference between the S-based ILS and the FE-based ILS is not very meaningful for the small and medium instances, whereas the gap grows to become unbearable and in fa-vor of the S-based ILS version for the large-sized instances. As a result, we use the S_ILS version to perform numerical experiments. The preliminary numerical experiments show that introducing worker well-being in order picking does not significantly affect overall system performance (i.e. number of references picked per day). In other words, our results support the idea of placing the well-being of the industry worker at the center of the production process, as suggested by Industry 5.0.