Data-Driven Decision-Making in Sequential Systems Under Uncertainty and Imperfect Information

Abstract

This dissertation examines decision-making in sequential systems when information is in- complete, uncertain, or costly to obtain. The first two contributions study inspection allocation in genetic manufacturing systems, where undetected nonconformities increase total cost and reduce process efficiency. Contribution 1 develops a Markov decision process model for a two- stage system consisting of polymerase chain reaction and cloning, with sequencing as the final inspection stage. Computational results identify stable policy patterns across a broad range of parameter settings. In many scenarios, the preferred strategy is to skip inspection after poly- merase chain reaction and use capillary electrophoresis after cloning. The results also show that at least one inspection before final sequencing is usually justified and that, when sequencing becomes sufficiently inexpensive, the best policy is to use sequencing after polymerase chain reaction and no inspection after cloning. Contribution 2 extends the framework to a system with two parallel synthesis lines, Gibson assembly, and final sequencing. The findings show that the best inspection strategy depends on where defects can occur, how they combine during assem- bly, and how much cost is incurred if defective fragments continue to later stages. Early in- spections are often valuable because they prevent additional processing of defective fragments, whereas high-accuracy inspections become more attractive at later stages when missed defects are more costly. Contribution 3 addresses planned missingness in an Internet of Things temper- ature sensor stream collected under an energy-saving sampling policy. A one-minute series is reconstructed from ten-minute observations using Gaussian Process and Piecewise Cubic Her- mite Interpolating Polynomial imputation and then evaluated with XGBoost, Fourier-enhanced XGBoost, Prophet, and a novel hybrid forecasting model. The proposed hybrid model achieves the best forecasting accuracy under both imputations, Prophet ranks second, and Gaussian Pro- cess imputation performs better overall. Together, these contributions show that better decisions in sequential systems depend on strategically acquiring and reconstructing information.