Mutation-Augmented NEAT (M-NEAT): ImprovingNeuroevolutionary Performance in DivergentMulti-Echelon Inventory Systems

Managing complex Divergent Multi-Echelon Inventory Systems (DMEIS) with traditional policies is challenging. This research advances the paradigm of using Reinforcement Learning (RL) by applying and enhancing the Neuroevolution of Augmenting Topologies (NEAT) algorithm for this task, where standard NEAT is often hindered by evolutionary stagnation. To address this, the paper introduces Mutation-AugmentedNEAT (M-NEAT), a novel variant that introduces Alternate Crossover Limitation and Alternative Mutation Probability to modify genetic operators, enabling concurrent structural evolution in less dominant lineages to enhance population diversity. A two-phase numerical study was conducted: first, a hyperparameter evaluation selected themost effective configurations for bothNEAT andM-NEAT. Second, a comparative analysis benchmarked the baseline performance againstProximalPolicyOptimization(PPO) andthen evaluatedM-NEAT againststandard NEAT across various operational scenarios. Experimental results show M-NEAT mitigates stagnation (reducing dominance streaks by 64.06%) and achieves a 12.83% total cost reduction over the benchmark, significantly outperforming standard NEAT. The analysis also reveals that M-NEAT learns a more direct, monotonic strategy in response to increasing costs. This study validates M-NEAT as a robust neuroevolutionary approach for complex supply chain optimization. These findings demonstrate the robustness of M-NEAT and highlight the potential of neuroevolution-based reinforcement learning to derive robust control policies for complex multi-echelon supply networks where analytic solutions are infeasible.