Prescriptive modeling for counterfactual inferences

Abstract

In real-life scenarios, conducting experiments or simulations to optimize out-comes can be costly in terms of time and resources. This thesis explores the utilization of trained neural networks for predictive modeling and optimization to address this challenge. The methodology involves training neural networks on historical data or simulated environments to capture complex relationships be-tween input variables and outputs. We then employ optimization techniques to explore parameter/input spaces and identify optimal configurations for desired outputs. Importantly, this approach enables us to conduct counterfactual analyses, allowing us to assess how changes in input parameters would affect outputs. We present case studies utilizing two distinct real-life scenarios: firstly, the public simulation model FluTE, where we demonstrate the effectiveness of our approach in optimizing strategies to alleviate the spread of infectious diseases. Secondly, we tackle an assortment problem and demonstrate how decision-making processes in retail settings can be assisted by trained neural networks to maximize profitability. We then also suggest an improved methodology to control the uncertainty in predicted outputs from neural network. We utilize dropout networks to quantify variability in the output predictions and embed them into the optimization model. Computational experiments are conducted with the two case studies and customized problem specific methodologies are suggested that includes decomposition methods and heuristics.