Loading...

Supply Chain Optimization with Perishable Products Through Demand Forecasting by a Reinforcement Learning Algorithm

Shams Shemirani, Sadaf | 2023

35 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56412 (01)
  4. University: Sharif University of Technology
  5. Department: Industrial Engineering
  6. Advisor(s): Khedmati, Majid
  7. Abstract:
  8. Using an efficient method to manage inventory systems is always a challenging issue in supply chain optimization. In supply chains including perishable goods, it is possible to reduce waste and other costs by identifying uncertain demand patterns and managing inventory levels at different stages of the supply chain. Considering the uncertainty and complex conditions of supply chains in the real world, in order to create a suitable model to express these conditions, various uncertain factors must be considered, each of which affects the supply chain inventory level in some way. In this research, a multi-level perishable supply chain model with uncertain demand, lead time and deterioration rate is represented. This study is done in three main parts; In the first part, the supply chain model with real-wold assumptions is created; In the second part, the mathematical model of the supply chain is converted into a Markovian stochastic process, and in the last part, a deep reinforcement learning network suitable for solving the model of the problem is presented, respectively. The supply chain includes four stages of producer, distributor, retailer and customer. For better management of warehouse inventory, inventory on the way between members and backlog inventory, multi-agent reinforcement learning has been used, by which, the order quantity of the members of each stage to their upstream stages is determined through training the reinforcement learning algorithm on the problem environment in order to maximize the profit of the entire supply chain, reduce the effect of leather whip and the failure of products in the chain. Furthermore; Using this approach, two different policies are compared according to the design of the supply chain model, in one of which, for each unsatisfied demand, sales are lost, and in the other case, backlogs are considered. In order to choose the appropriate reinforcement learning algorithm to solve the problem, three algorithms PPO, DQN and a traditional replenishment policy are compared on a two-stage supply chain with uncertain demand, and the result can be considered and generalized for the presented model. As the result of this comparison, the PPO algorithm has shown better performance than the two other algorithms. Finally, this algorithm has been used in a case study, and the results show that the lost sales condition is more suitable than backlogs due to the characteristics of the model
  9. Keywords:
  10. Inventory Management ; Uncertain Demand ; Deep Reinforcement Learning ; Multi-level Supply Chain ; Markov Decision Making ; Multi-Agent Reinforcement Learning ; Perishable Supply Chain

 Digital Object List

 Bookmark

No TOC