A predictive analysis of the Department of Defense distribution system utilizing random forests

Abstract

This thesis develops machine-learning models capable of predicting Department of Defense distribution system performance of United States Marine Corps ocean requisitions to the United States Pacific Command area of operations. We use historical data to develop a model for each sub-segment of the Transporter leg within the distribution pipeline and develop two different models to predict the ocean transit sub-segment based on Hawaii and non-Hawaii destinations. We develop a linear regression, regression tree and random forest model for each sub-segment and find that the weekday and month in which requisitions begin the Transporter segment are among the most significant drivers in variability. United States Transportation Command currently uses the average performance per sub-segment to estimate Transporter length, and our models, when applied to the test set, perform considerably better than the average. We conclude that the random forest models provide the best and most robust results for most sub-segments. However, we encounter several issues concerning missing values within our dataset, which we suspect artificially inflate the significance of some of our predictor variables. We recommend refining data collection processes in order to collect observations that are more accurate and applying the same methodologies in the future.