Many modern industrial processes demand automation, opening many opportunities for integrating a robot-based fulfillment system. This work delves into the collaborative nature of the Robotic Mobile Fulfillment System (RMFS). In this complex environment, robots deliver products to humans who fulfill orders. The RMFS is a complex problem that integrates various optimization subproblems. We focus on optimizing the Pod Allocation Problem (PAP), a crucial part of the RMFS, by following an algorithm selection approach. Our proposal implements neural networks, specifically Multi-Layer Perceptrons (MLP) models, to map the problem features into a suitable solver for the PAP. The way we achieve this requires analyzing the behavior of the available solvers on different training instances, properly characterized using straightforward features. Then, a Multi-Layer Perceptron learns to relate the features that characterize the instances to one suitable solver, given their historical performance. Once the MLP has learned this pattern, it is ready to recommend the most suitable solver for new cases. The result is a method capable of adapting to distinct scenarios, thereby increasing its generalization capabilities. This work analyzes 272 RMFS instances, which we solved using six Pod Allocation Solvers and the algorithm selectors produced through our approach. Our proposed architectures proved competent in terms of two different performance metrics: throughput time and the number of processed orders. Moreover, the algorithm selectors produced can rival and even outperform the best low-level solvers under some particular conditions.
Benavides-Robles et al. (Sun,) studied this question.