Author Type

Graduate Student

Date of Award

Summer 6-19-2026

Document Type

Thesis

Publication Status

Version of Record

Submission Date

July 2026

Department

Computer and Electrical Engineering and Computer Science

College Granting Degree

College of Engineering and Computer Science

Department Granting Degree

Electrical Engineering and Computer Science

Degree Name

Master of Science (MS)

Thesis/Dissertation Advisor [Chair]

Zhen Ni,

Thesis/Dissertation Co-Chair

Evangelos I. Kaisar

Abstract

Reinforcement learning methods have demonstrated strong performance on vehicle routing benchmarks, yet their behavior under severe capacity constraints remains unexplored. This dissertation investigates whether PPO-based neural routing policies maintain operational viability when vehicle capacity is severely constrained, as occurs in resource-limited rural logistics settings.

Through controlled experiments on synthetic instances and validation on real-world rural healthcare networks in Florida, this research reveals a critical capacity threshold effect. Moderate capacity reductions from 40 to 20 produce negligible performance loss (4.1%), while severe reductions to capacity 10 trigger catastrophic failure with 243% degradation, manifested through degenerate single-customer routing patterns. Convergence analysis identifies the underlying mechanism: low-capacity configurations exhibit chronic critic instability (loss 14.77 versus 0.68) due to extended decision horizons that fundamentally compromise the actor-critic learning dynamics.

Real-world validation across healthcare facility networks of varying scale (20, 50, 100 nodes) confirms that this threshold generalizes beyond synthetic distributions, with degradation intensifying systematically as problem size increases. The findings establish an operational feasibility boundary between capacities 10 and 20, below which PPO-based routing becomes deployment-infeasible despite producing formally valid solutions.

These results demonstrate that benchmark performance does not guarantee robustness under operational constraints, with immediate implications for deploying learning-based routing in rural logistics contexts where vehicle capacity is dictated by infrastructure limitations rather than optimization objectives.

Share

COinS