Work

Evolutionary-assisted multi-agent reinforcement learning for waste electrical and electronic equipment inventory optimization under uncertainty

Year: 2026

Type: article

Abstract: Waste electrical and electronic equipment (WEEE) inventory optimization has attracted increasing attention with the growing emphasis on circular economy and sustainable development concepts. Effective uncertainty management is crucial for inventory optimization because recycling quantities fluctuate and custom order demands can arise unexpectedly. Although single-agent reinforcement learning methods can learn control policies for sequential decision-making tasks, they neglect collaboration and information sharing among multiple entities, making the resulting solutions unreliable in practical applications. In this paper, an evolutionary-assisted multi-agent reinforcement learning (EAMARL) method is proposed to achieve multi-period, multi-echelon WEEE supply chain inventory optimization. Specifically, a control model is proposed for WEEE inventory optimization under uncertain recycling quantities and custom order demands. The model integrates make-to-order (MTO) and make-to-stock (MTS) production modes, and the resulting optimization problem is formulated as a decentralized Markov decision process. A multi-agent reinforcement learning (MARL) algorithm interacts with the WEEE supply chain environment to train control policies that maximize total profit. To address unstable convergence and limited exploration in MARL approaches, EAMARL integrates a population-based evolutionary algorithm that provides diverse exploratory experiences to support the training of MARL agents. Experiments in open-loop and closed-loop WEEE supply chain networks show that EAMARL improves average profit over the best-performing baseline by 12.09% and 14.39%, respectively. The normalized area under the learning curve of EAMARL indicates high learning efficiency and favorable convergence performance, while boxplot analysis demonstrates greater performance stability. Holm-adjusted Wilcoxon signed-rank tests and paired 95% confidence intervals further provide statistical support for the performance improvements.

Source: Engineering Applications of Artificial Intelligence

DOI: 10.1016/j.engappai.2026.116478