Inventory planning of perishable items using reinforcement learning
Loading...
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2023.
Abstract
Managing perishable inventory effectively is vital in diverse sectors like grocery, pharmaceuticals, composite materials, agriculture, blood. The challenge lies in reducing costs while handling items with limited shelf lives and changing demand. This study delves into the potential of computational techniques, notably the reinforcement learning methods Q- learning and SARSA, to tackle this intricate issue. We turned to reinforcement learning because traditional approaches struggle with increased complexity as the problem grows. We began our exploration with backward dynamic programming for a basic perishable inventory model. This model covered 10 periods and 3 age classes under a deterministic demand. We then expanded this framework to address more unpredictable demand patterns, both stationary and non-stationary, crafting optimal value functions for each. Our study also ventured into the Q- value approach, where transition probabilities were predefined, comparing the results to traditional dynamic programming. We further evaluated Q-learning and SARSA to see how close they converge to optimal. As the problem’s complexity rose, especially with advanced demand scenarios like Advance Demand Information and more age classes beyond three, traditional methods fell short. In contrast, reinforcement learning proved nimble, especially in tackling more intricate inventory challenges. Our findings underline that reinforcement learning methods can approximate the near-optimal results achieved by dynamic programming in simpler scenarios. More remarkably, as the problem’s intricacy grew, reinforcement learning continued to offer solutions, suggesting its promise in addressing even more complex inventory challenges.