Data-Driven Fleet Lifecycle Management: Integrating Interpretable Machine Learning and Stochastic Optimization
Date Issued
May 2026
Author(s)
Phillips, Lanette Marie
The University of Tennessee
Advisor(s)
Andrew Yu
Additional Advisor(s)
Qiang He
Jim Ostrowski
Beau Schelble
Abstract
Public agencies manage large vehicle fleets under strict budget constraints, service-level requirements, and public accountability. Replacement decisions are often governed by uniform age or mileage rules that simplify administration but fail to reflect heterogeneous utilization, nonlinear cost behavior, and uncertainty in maintenance and downtime. This dissertation develops an integrated analytical framework that combines stochastic optimization, interpretable machine learning, and decision-focused integration to support transparent and economically defensible fleet planning. First, fleet replacement and capacity planning are formulated as a two-stage stochastic program capturing the commitment–recourse trade-off between owned capacity and adaptive operational decisions under uncertainty. Decomposition methods—including Benders decomposition, Lagrangian relaxation, and Progressive Hedging—are evaluated to ensure tractability and policy consistency. Second, a heterogeneity-aware predictive model of annual operating cost is developed using variance-stabilized gradient boosting trained on log-transformed expenditures. Under VIN-safe and agency-blocked validation that prevents temporal and organizational leakage, the selected model demonstrates strong out-of-sample accuracy and calibration despite heavy-tailed cost distributions. SHAP-based explanations reveal nonlinear age effects, ownership–utilization interactions, and agency-level structural variation. Counterfactual analysis shows no stable lease–own crossover within the empirically supported mid-life policy window once censoring and compositional effects are controlled. Finally, predictive cost structure is embedded within the stochastic model through a Decision–Coupled Learning framework. Model performance is evaluated by downstream economic impact—expected cost, risk exposure, and policy regret—rather than statistical error alone. The resulting framework provides a transparent, uncertainty-aware, and operationally implementable decision infrastructure for public-sector fleet management.
Degree
Doctor of Philosophy
Major
Industrial Engineering
File(s)![Thumbnail Image]()
Name
PHILLIPS-PRIMARY-2026.pdf
Size
3.85 MB
Format
Adobe PDF
Checksum (MD5)
eabe191e49bf71d6000762f212926435