Bilevel Stochastic Optimization in Decentralized Assemble-To-Order Supply Chain

The Assemble-to-Order (ATO) problem is a production and inventory management challenge commonly encoun- tered in industries where products are assembled from a variety of components or subassemblies. In an ATO environment, a company holds a stock of components and assembles final products only when customer orders are received. This approach allows for a balance between customization (since products are assembled based on spe- cific customer requirements) and efficiency (since the company does not need to stock large quantities of finished goods). There has been a rich source of literature regarding the problem that deals with a variety of versions of ATO model. However, the main focus of the main stream literature is to prove theoretically a number of certain properties of the model and design control policies to optimize predefined metrics. This approach isolates the assembler as the sole rational decision maker of the process and ignore the complexity and the contractual nature of supply chain coordination. As mentioned in Kok, de and Graves [2003], it is impossible for the supplier to force the assembler to receive more components than the ordered amount but in return, there is not a strict compliance regime to force the supplier to deliver the desirable amount of components to the assembler due to objective and subjective factor. Therefore, it is necessary to create incentive by sharing information on the uncertainty of demand so that each actor in the supply chain to behave in a predictable way. There are a few similar papers regarding this line of research. One of the is Gerchak and Wang [2004] that investigate a one-product model with revenue-sharing contract which is showed to yield higher system profit. The other is Bernstein et al. [2007] in which the authors investigate a 2-product, 3-components (an M system) that shows possible capacity imbalance and inefficiency cause by decentralized decision making. Our interest is to build a large scale bilevel stochastic model to handle the size of a realistic instance that optimizes the profit of the assembler.