Improving scholarship assignment using approximate dynamic programming: A Chilean case study

Sofía Redondo

doi:10.1016/j.seps.2025.102296

Return

Improving scholarship assignment using approximate dynamic programming: A Chilean case study

2025-08-09

PRE

DOI：10.1016/j.seps.2025.102296

Original

Original request for help

Save

Abstract

En 中文

• Scholarship assignment framed as an MDP solved via ADP The annual allocation problem, subject to budget limits and renewal uncertainty, is modeled as a Markov Decision Process. An Approximate Dynamic Programming (ADP) scheme (affine value function approximation) with column generation overcomes the curse of dimensionality. • 7% to 9% increase in beneficiaries at constant budget Simulations calibrated to the Chilean Beca Presidente de la Rep’ublica show that the ADP policy could award 7% to 9% more scholarships per year without additional spending when compared with current administrative rules. • Stochastic renewal length explicitly guaranteed. When a new scholarship is granted, the model ensures future funding for all renewals by treating scholarship duration as a random variable driven by each student ′s academic trajectory. • Empirical case study with 2014 to 2020 administrative records. Six annual renewal files (78 variables) and six application files (51 variables) are cleaned, merged into cohorts, and used both to calibrate the model and to benchmark six alternative policies. • Data driven grouping and transition dynamics. Applicants and recipients are clustered by education level and socioeconomic attributes; transition probabilities are estimated from historical shares, while new arrivals follow a Poisson process with the empirical mean rate. • Robust long-run performance. Over 50 replications of a 650-year horizon, the ADP policy consistently outperforms six benchmark strategies on seven key metrics-granting 4% to 13% more scholarships, lowering cut- off scores, and improving budget utilization.

Keywords:

Scholarship allocation

Markov Decision Process

Approximate Dynamic Programming

Budget optimization

Stochastic renewal modeling