Employing Neuroevolution of Augmenting Topologies (NEAT) in Linear Multi-Echelon Inventory Systems

Abstract

Reinforcement learning has emerged as a leading algorithmic approach due to its successful applications across various domains. While many implementations favour the model-free approach for its aptitude for handling complex problems, its learning curve tends to be slower. Given the intricacies of the Linear Multi-Echelon Inventory System, a model-based approach might be more fitting, offering faster learning rates. This study seeks to integrate Neuroevolution of Augment Topologies (NEAT) – a hybrid of model-based reinforcement learning and evolutionary algorithms – into such an inventory system. Furthermore, the research delves into hyperparameter tuning, experimenting with seven specific hyperparameters to discern the most efficient combination and understand their interplay. Benchmarking against the model-free Proximal Policy Optimisation (PPO) serves as a measure of NEAT's effectiveness. Findings indicate that when optimally tuned, NEAT can slash total costs by 25.02% compared to PPO. Impressively, NEAT achieves this peak performance in a mere 1,000 generations, significantly outpacing PPO's learning trajectory.