Fairness-Constrained Distributionally Robust Causal Policy Optimization for Adaptive Instruction in Higher Education

Adaptive instructional policies in higher education often optimize short-term engagement at the expense of durable learning gains, and they frequently neglect heterogeneity across student subpopulations. We propose a fairness-constrained distributionally robust causal policy optimization framework that replaces standard deep reinforcement learning agents with a causally grounded decision-making pipeline. A structural causal model of the educational environment is constructed first, and doubly robust estimation is then applied to observational log data so that confounding bias in the counterfactual evaluation of instructional actions is mitigated. To guard against distribution shifts and adversarial perturbations in student populations, the policy optimization problem is formulated as a distributionally robust objective that maximizes worst-case expected utility over an ambiguity set centered on the empirical distribution. The policy itself is represented by reduced-rank policy trees, which preserve interpretability while remaining computationally tractable, and multi-objective Bayesian optimization is employed to balance average performance, cross-subgroup variance, and explicit disparity constraints. Fairness is enforced through Lagrangian relaxation, while causal effect decomposition separates direct instructional effects from those mediated by motivation or cognitive load. Periodic off-policy evaluation and counterfactual subgroup audits provide closed-loop validation, triggering retraining or constraint tightening when inequities emerge. The main contribution is an integrated methodology that unites causal inference, distributional robustness, and subgroup fairness within a single policy learning loop for higher education. This approach is significant because it offers a principled path toward instructional policies that remain effective, equitable, and auditable throughout their deployment lifecycle.