A constrained multi-item EOQ inventory model for reusable items: Reinforcement learning-based differential evolution and particle swarm optimization

Abstract

The growing environmental concerns, governmental regulations, and significant cost savings are the primary motivations for companies to consider the reuse and recovery of products in their inventory system. The previous research ignored several realistic features of reusable items inventory systems, such as the presence of multiple products and operational constraints. For the first time, this paper presents a new multiproduct economic order quantity inventory model for an inventory system of reusable products. The goal of the model is to determine the optimal replenishment quantity and reuse quantity of each item so that the system's total cost is minimized. Several operational constraints are considered to provide a more realistic framework, such as the total available budget, warehouse space, and holding cost. Due to the nonlinearity of the presented model, differential evolution (DE) and particle swarm optimization (PSO) algorithms are utilized as two solution approaches. However, these algorithms' performance is highly dependent on their control parameters. Therefore, for the first time, two new variants of these algorithms, called DEQL and PSOQL, are presented in this study, where the control parameters of algorithms are not pre-determined. A powerful reinforcement learning algorithm, Q-learning, adapts these values intelligently. In other words, as a research contribution, this research aims at introducing a new variant of hybrid the DE and PSO algorithms in which a machine learning algorithm controls the value of metaheuristic parameters. The other parameters of the proposed algorithms are tunned employing the Taguchi method. Extensive numerical examples are established in different sizes, and the outputs are discussed in terms of several criteria. Statistical analysis of the results is performed, demonstrating that the proposed reinforcement learning-based parameter adaption has significantly improved algorithms' performance.