A Sustainable Deep Reinforcement Learning Framework for Personalized Learning: Balancing Learning Gains, Resource Efficiency, and Educational Equity: A Simulation Study

This study is a fully synthetic simulation study; no real-student data are used. Personalized learning systems must balance achievement, resource use, and equity. We examine a deep reinforcement learning formulation motivated by Sustainable Development Goal 4. A shared agent selects among six instructional actions for 500 synthetic students over 80 timesteps. The student model combines difficulty matching, engagement-dependent learning, and forgetting. A five-term reward incorporates learning gain, engagement, time cost, resource consumption, and remedial fairness. In the main comparison, a Deep Q-Network (DQN) and a Proximal Policy Optimization agent trained for 200 episodes are evaluated against tabular Q-learning, a rule-based policy, and random selection over 20 episodes. DQN attains the highest composite reward (2.428) and the lowest resource use (56.21 simulation units per student). Its normalized gain (0.685) is slightly below that of random selection (0.704), while its gain and equity per unit of resource are the highest (0.0122 and 0.0141); random selection lifts initially low-performing students most. Because the reward explicitly penalizes resource use and the learned policy concentrates on low-cost actions, the composite-reward ranking should be read together with the underlying metrics rather than as an independent achievement advantage. The fairness term optimizes a bottom-quartile-to-median gap reduction, which is distinct from the dispersion-based equity index used in evaluation; the two measures are reported side by side and are not used interchangeably. Further analyses vary the reward, the learner model, and the training budget, and show that the rankings shift with the objective and with the simulated learning dynamics. A robustness block conducted with a documented re-implementation—the primary experimental record of this study—adds independent five-seed training runs, stronger static comparators, and reward-definition controls. Under repeated training, the three learned policies are descriptively close on learning gain, so the original single-run DQN–PPO separation is not reproduced; a myopic cost-aware greedy comparator that exploits the same reward information without learning attains higher gain, lower resource use, and higher equity than every learned policy. In this calibrated simulator the sequential reinforcement learning component therefore has no demonstrated advantage over short-sighted exploitation of the same objective. The study’s contribution is accordingly the formulation itself—making the learning, engagement, time, resource, and fairness trade-offs explicit and measurable—together with a reproducible diagnostic record of how strongly the conclusions depend on the reward design and the simulated learner dynamics; effectiveness with real students remains untested.

Authors

Institutions

Publication Details

Journal
Sustainability
Published
2026-11-12
DOI
https://doi.org/10.3390/su182211524
Primary Topic
Intelligent Tutoring Systems and Adaptive Learning
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Sustainable Deep Reinforcement Learning Framework for Personalized Learning: Balancing Learning Gains, Resource Efficiency, and Educational Equity: A Simulation Study

Mohamad Nizam Nazarudin, Norzaini Azman, Xie Xiaoning
Sustainability
Intelligent Tutoring Systems and Adaptive Learning
article

A Sustainable Deep Reinforcement Learning Framework for Personalized Learning: Balancing Learning Gains, Resource Efficiency, and Educational Equity: A Simulation Study

Mohamad Nizam Nazarudin, Norzaini Azman, Xie Xiaoning
article en

Abstract

This study is a fully synthetic simulation study; no real-student data are used. Personalized learning systems must balance achievement, resource use, and equity. We examine a deep reinforcement learning formulation motivated by Sustainable Development Goal 4. A shared agent selects among six instructional actions for 500 synthetic students over 80 timesteps. The student model combines difficulty matching, engagement-dependent learning, and forgetting. A five-term reward incorporates learning gain, engagement, time cost, resource consumption, and remedial fairness. In the main comparison, a Deep Q-Network (DQN) and a Proximal Policy Optimization agent trained for 200 episodes are evaluated against tabular Q-learning, a rule-based policy, and random selection over 20 episodes. DQN attains the highest composite reward (2.428) and the lowest resource use (56.21 simulation units per student). Its normalized gain (0.685) is slightly below that of random selection (0.704), while its gain and equity per unit of resource are the highest (0.0122 and 0.0141); random selection lifts initially low-performing students most. Because the reward explicitly penalizes resource use and the learned policy concentrates on low-cost actions, the composite-reward ranking should be read together with the underlying metrics rather than as an independent achievement advantage. The fairness term optimizes a bottom-quartile-to-median gap reduction, which is distinct from the dispersion-based equity index used in evaluation; the two measures are reported side by side and are not used interchangeably. Further analyses vary the reward, the learner model, and the training budget, and show that the rankings shift with the objective and with the simulated learning dynamics. A robustness block conducted with a documented re-implementation—the primary experimental record of this study—adds independent five-seed training runs, stronger static comparators, and reward-definition controls. Under repeated training, the three learned policies are descriptively close on learning gain, so the original single-run DQN–PPO separation is not reproduced; a myopic cost-aware greedy comparator that exploits the same reward information without learning attains higher gain, lower resource use, and higher equity than every learned policy. In this calibrated simulator the sequential reinforcement learning component therefore has no demonstrated advantage over short-sighted exploitation of the same objective. The study’s contribution is accordingly the formulation itself—making the learning, engagement, time, resource, and fairness trade-offs explicit and measurable—together with a reproducible diagnostic record of how strongly the conclusions depend on the reward design and the simulated learner dynamics; effectiveness with real students remains untested.

SustainabilityVol. 18(22)
National University of Malaysia (MY)
Quality education, Reduced inequalities
Intelligent Tutoring Systems and Adaptive Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon