Data-Driven Decision Making in Operations Management
Encouraged by the plethora of advances in artificial intelligence (AI) in the past decade, this thesis studies the length to which we can push various business operations with new technologies, in our theoretical understanding and practical performance alike. Towards this goal, this thesis develops...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2023
|
| Online Access: | https://hdl.handle.net/1721.1/151589 |
| _version_ | 1826194346324525056 |
|---|---|
| author | Gong, Xiaoyue |
| author2 | Simchi-Levi, David |
| author_facet | Simchi-Levi, David Gong, Xiaoyue |
| author_sort | Gong, Xiaoyue |
| collection | MIT |
| description | Encouraged by the plethora of advances in artificial intelligence (AI) in the past decade, this thesis studies the length to which we can push various business operations with new technologies, in our theoretical understanding and practical performance alike. Towards this goal, this thesis develops data-driven decision-making methods for a selection of challenging emerging problems in supply chain and other business operations.
In the first module of the thesis (Chapters 2 and 3), we invent reinforcement learning methods with provable optimality guarantees for inventory management problems. The challenge in the inventory problems that we are interested in is that the demand distribution varies over time according to some natural cyclic patterns (such as weekly sales cycles), and we are in the online setting where we do not have prior knowledge of the demand distribution or access to prior data. Solutions to these inventory models have been carefully studied for decades in the offline setting where the cyclic demand distribution is known beforehand; however, very few results have been attained in the online setting. The complexity of the problem motivated us to introduce reinforcement learning into the picture. Our design of a reinforcement learning algorithm has an optimal regret bound for a number of inventory models with unknown cyclic demands that we study in these chapters.
In the second module of the thesis (Chapter 4), we study online assortment optimization for reusable resources. E-commerce platforms like Amazon and Expedia constantly endeavor to recommend more favorable assortments of products and services to their customers. The choice of assortment influences customer purchasing decisions, and can thus significantly impact the platform’s revenue. We consider assortment optimization with reusable resources, which means that the product returns to the inventory once the customer has finished using it. Reusability arises in major applications including cloud services, physical storage, and make-to-order service. The unpredictability of the usage times means that planning ahead becomes more challenging. We show that a simple greedy policy is 1/2 competitive for online assortment optimization with reusable resources. This means that on average, the greedy policy earns at least half the revenue of a clairvoyant optimal policy which has access to much more information. This result is surprising because the greedy policy does not take into account the customer or usage time distributions, both of which are necessary to solve for the optimal policy.
In the third module of the thesis (Chapter 5), we develop practical solutions for the cloud service supply chain at Microsoft Azure. The cloud computing industry boomed in the past few years as digitization continues to take place globally and as remote work becomes more of a norm. A main challenge faced by cloud service providers is to deploy cloud server hardware under demand uncertainty, without incurring unnecessarily large operational costs. We formulate the underlying optimization problem as a two-stage stochastic program. We then develop exact Benders-type algorithms that exploit the structure of the second stage problem. We test our proposed algorithms with numerical experiments based on real production traces from Microsoft Azure, which demonstrate noticeable advantages of our algorithms over existing heuristics used in production. Given the large scale of the problem, our deployment policy could potentially lead to savings of hundreds of millions of dollars per year. |
| first_indexed | 2024-09-23T09:54:36Z |
| format | Thesis |
| id | mit-1721.1/151589 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T09:54:36Z |
| publishDate | 2023 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1515892023-08-01T04:08:13Z Data-Driven Decision Making in Operations Management Gong, Xiaoyue Simchi-Levi, David Massachusetts Institute of Technology. Operations Research Center Encouraged by the plethora of advances in artificial intelligence (AI) in the past decade, this thesis studies the length to which we can push various business operations with new technologies, in our theoretical understanding and practical performance alike. Towards this goal, this thesis develops data-driven decision-making methods for a selection of challenging emerging problems in supply chain and other business operations. In the first module of the thesis (Chapters 2 and 3), we invent reinforcement learning methods with provable optimality guarantees for inventory management problems. The challenge in the inventory problems that we are interested in is that the demand distribution varies over time according to some natural cyclic patterns (such as weekly sales cycles), and we are in the online setting where we do not have prior knowledge of the demand distribution or access to prior data. Solutions to these inventory models have been carefully studied for decades in the offline setting where the cyclic demand distribution is known beforehand; however, very few results have been attained in the online setting. The complexity of the problem motivated us to introduce reinforcement learning into the picture. Our design of a reinforcement learning algorithm has an optimal regret bound for a number of inventory models with unknown cyclic demands that we study in these chapters. In the second module of the thesis (Chapter 4), we study online assortment optimization for reusable resources. E-commerce platforms like Amazon and Expedia constantly endeavor to recommend more favorable assortments of products and services to their customers. The choice of assortment influences customer purchasing decisions, and can thus significantly impact the platform’s revenue. We consider assortment optimization with reusable resources, which means that the product returns to the inventory once the customer has finished using it. Reusability arises in major applications including cloud services, physical storage, and make-to-order service. The unpredictability of the usage times means that planning ahead becomes more challenging. We show that a simple greedy policy is 1/2 competitive for online assortment optimization with reusable resources. This means that on average, the greedy policy earns at least half the revenue of a clairvoyant optimal policy which has access to much more information. This result is surprising because the greedy policy does not take into account the customer or usage time distributions, both of which are necessary to solve for the optimal policy. In the third module of the thesis (Chapter 5), we develop practical solutions for the cloud service supply chain at Microsoft Azure. The cloud computing industry boomed in the past few years as digitization continues to take place globally and as remote work becomes more of a norm. A main challenge faced by cloud service providers is to deploy cloud server hardware under demand uncertainty, without incurring unnecessarily large operational costs. We formulate the underlying optimization problem as a two-stage stochastic program. We then develop exact Benders-type algorithms that exploit the structure of the second stage problem. We test our proposed algorithms with numerical experiments based on real production traces from Microsoft Azure, which demonstrate noticeable advantages of our algorithms over existing heuristics used in production. Given the large scale of the problem, our deployment policy could potentially lead to savings of hundreds of millions of dollars per year. Ph.D. 2023-07-31T19:50:52Z 2023-07-31T19:50:52Z 2023-06 2023-07-13T16:03:40.547Z Thesis https://hdl.handle.net/1721.1/151589 In Copyright - Educational Use Permitted Copyright retained by author(s) https://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Gong, Xiaoyue Data-Driven Decision Making in Operations Management |
| title | Data-Driven Decision Making in Operations Management |
| title_full | Data-Driven Decision Making in Operations Management |
| title_fullStr | Data-Driven Decision Making in Operations Management |
| title_full_unstemmed | Data-Driven Decision Making in Operations Management |
| title_short | Data-Driven Decision Making in Operations Management |
| title_sort | data driven decision making in operations management |
| url | https://hdl.handle.net/1721.1/151589 |
| work_keys_str_mv | AT gongxiaoyue datadrivendecisionmakinginoperationsmanagement |