Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Department of Operational Sciences

First Advisor

Matthew J. Robbins, PhD


Effective personnel management policies in the United States Air Force (USAF) require methods to predict the number of personnel who will remain in the USAF as well as to replenish personnel with different skillsets over time as they depart. To improve retention predictions, we develop and test traditional random forest models and feedforward neural networks as well as partially autoregressive forms of both, outperforming the benchmark on a test dataset by 62.8% and 34.8% for the neural network and the partially autoregressive neural network, respectively. We formulate the workforce replenishment problem as a Markov decision process for active duty enlisted personnel, then extend this formulation to include the Air Force Reserve and Air National Guard. We develop and test an adaptation of the Concave Adaptive Value Estimation (CAVE) algorithm and a parameterized Deep Q-Network on the active duty problem instance with 7050 dimensions, finding that CAVE reduces costs from the benchmark policy by 29.76% and 17.38% for the two cost functions tested. We test CAVE across a range of hyperparameters for the larger intercomponent problem instance with 21,240 dimensions, reducing costs by 23.06% from the benchmark, then develop the Stochastic Use of Perturbations to Enhance Robustness of CAVE (SUPERCAVE) algorithm, reducing costs by another 0.67%. Resulting algorithms and methods are directly applicable to contemporary USAF personnel business practices and enable more accurate, less time-intensive, cogent, and data-informed policy targets for current processes.

AFIT Designator


DTIC Accession Number