💻 computer science

A genetic algorithm for student academic resource allocation

This paper proposes a Genetic Algorithm with a specialized constraint repair mechanism to solve the NP-hard problem of optimally allocating mathematics learning resources to high school students under strict time constraints, demonstrating fast convergence and high solution quality for real-time educational decision support.

Original authors: Ana F. Hernández, Andrej Franulic, Fernando Jiménez

Published 2026-07-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ana F. Hernández, Andrej Franulic, Fernando Jiménez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a spaceship, but instead of stars, your cargo hold is filled with thousands of different tools, gadgets, and maps. Your mission? To help a single passenger reach their destination as quickly and happily as possible. The problem is, your passenger has a very specific set of needs: maybe they get lost easily, maybe they are a speed demon, or maybe they have a limited amount of fuel. You can't just throw everything at them; you have to pick the perfect combination of items that fits in the fuel tank while giving them the best chance of success. This is the daily challenge of modern education. Teachers have a massive library of lessons, videos, and exercises, but every student learns differently. Some need extra help with reading numbers, while others are ready to sprint ahead. Figuring out the exact right mix of materials for one student is like trying to solve a giant, impossible puzzle where the pieces keep changing shape. This is where computer science steps in with a clever trick called a "Genetic Algorithm." Think of it as a digital evolution lab. Instead of a human trying to guess the perfect list, the computer creates thousands of random lists, tests them, and then "breeds" the best ones together, mixing and matching them over and over again until it evolves a solution that is almost perfect. It's nature's way of solving math problems: survival of the fittest, but for homework assignments.

This paper, written by a team of researchers from Spain, tackles exactly this puzzle: how to automatically pick the best math resources for a high school student. They frame the problem as a "0–1 binary combinatorial optimization" challenge. In plain English, this just means deciding for every single item in a catalog whether to include it (1) or leave it out (0), all while staying under a strict time limit. The researchers know that as the list of resources grows, trying to check every single possible combination becomes impossible for a computer to do quickly—like trying to count every grain of sand on a beach one by one. Because of this, they propose using their "Genetic Algorithm" to find a great solution without checking every single possibility.

The team built a system that first looks at a student's profile. They consider the student's current grade, how well they do in different math areas (like algebra or geometry), and even specific learning challenges. For example, they use a "diagnostic factor" to adjust the plan: a student with high abilities gets a factor of 1.0, while a student with dyslexia gets 0.7, and a student with dyscalculia (trouble with numbers) gets 0.5. This factor acts like a volume knob, turning down the intensity of the tasks to match the student's needs. The system also filters out resources that are too easy or too hard, keeping only the ones that match the student's current level and grade.

Once the system has a list of suitable resources, it runs its digital evolution. It starts by randomly picking a bunch of different combinations of resources. Then, it checks if any of these combinations break the rules—specifically, if the total time needed to finish all the selected resources exceeds the student's maximum allowed study time, which in their experiment was set to 8,100 minutes (or 135 hours). If a combination is too long, a special "repair mechanism" kicks in. It acts like a strict editor, randomly removing items from the list until the total time fits perfectly within the limit. This ensures that every solution the computer considers is actually possible to complete.

The researchers tested this system using a synthetic catalog of 1,000 math resources. They focused on a simulated student profile: a second-year high school student with dyslexia. After filtering out resources that didn't fit the student's level, the computer had 217 potential items to choose from. They ran the genetic algorithm 10 times, each time starting with a different random seed to see if the results were consistent. The results were impressive. The algorithm consistently found high-quality solutions, converging quickly to a "best fitness" score of 4.3333. In the best run, the system selected exactly 77 resources out of the 217 available, filling the student's schedule right up to the 8,100-minute limit without going over. The fact that the results were so similar across all 10 runs suggests the method is very stable and reliable.

The paper concludes that this approach works well for creating personalized learning plans in real-time. However, the authors are careful to note that this is a simulation based on a specific mathematical model. They don't claim to have solved education forever, but rather that they have a powerful tool for making these decisions. Looking ahead, the team suggests that the next step is to handle students with multiple challenges at once—like a student who is both gifted and has dyslexia. In that future scenario, the computer would have to balance competing goals, finding a "Pareto-optimal" solution where it does the best it can for all the student's different needs simultaneously. For now, though, this genetic algorithm stands as a promising way to turn a chaotic pile of educational materials into a tailored roadmap for a student's success.

Technical Summary: A Genetic Algorithm for Student Academic Resource Allocation

Problem Definition
This paper addresses the challenge of personalizing educational resource allocation for high school mathematics students within the Erasmus+ KA220-SCH project framework. The core problem is modeled as a 0–1 binary combinatorial optimization task. Given a finite catalog of pedagogical resources (categorized by subject area, difficulty, grade level, and type), the objective is to select an optimal subset for an individual student. This selection must maximize educational utility while adhering to a strict global constraint: the cumulative estimated execution time of selected resources cannot exceed a predefined maximum study time ( $T_{max}$ ).

The problem is identified as NP-hard due to its structural equivalence to the knapsack problem, where the search space grows exponentially ( $2^n$ ) with the number of available resources. Consequently, exact enumerative methods are deemed computationally intractable for real-time decision support in large-scale systems.

Methodology
To navigate the exponentially large search space, the authors propose a tailored Genetic Algorithm (GA) integrated with a specialized constraint handling mechanism. The methodology comprises the following components:

Mathematical Formulation: The objective function ( $f(x)$ $f (x)$ ) quantifies educational utility by weighting resources based on the student's academic profile. Key factors include:
- Area Need Weighting: Inversely scaling utility by the student's global evaluation score in a specific mathematical area, prioritizing weak subjects.
- Alignment Penalties: Adjusting utility based on the alignment of resource grade levels and difficulty with the student's current proficiency.
- Diagnostic Adjustment: Applying a scalar factor ( $s \in (0, 1]$ ) derived from expert-defined diagnoses (e.g., Dyslexia, ADHD, Dyscalculia) to modulate the utility gain, ensuring conservative task assignments for students with learning difficulties.
Genetic Algorithm Design:
- Representation: Solutions are encoded as binary vectors where a '1' indicates resource selection.
- Constraint Repair: A deterministic repair operator, RepairMaxTime, is employed to ensure feasibility. Instead of using penalty functions, this operator systematically deactivates randomly selected resources from infeasible solutions until the total time constraint is satisfied. This repair is applied post-initialization, post-crossover, and post-mutation.
- Evolutionary Operators: The algorithm utilizes Binary Tournament Selection for mating, Two-Point Crossover for recombination, and Bit-Flip Mutation for variation.
- Selection Strategy: An elitist $(\mu + \lambda)$ survival strategy is used, where the top $N$ individuals from the combined parent and offspring populations are retained for the next generation.

Experimental Evaluation
The proposed approach was implemented using the pymoo framework and evaluated on a synthetic dataset.

Setup: A master catalog of 1,000 resources was pre-filtered based on student grade and proficiency, resulting in an active decision space of 217 resources for a target student profile (Grade 2, diagnosed with Dyslexia). The time budget was set to 8,100 minutes.
Parameters: The GA was configured with a population size of 50, 300 generations, a crossover probability of 0.8, and a mutation probability of 0.1.
Runs: The experiment consisted of 10 independent runs with varying random seeds to assess stability.

Key Results

Convergence: The algorithm demonstrated rapid convergence, with significant fitness improvements occurring within the first 75 generations. The solution quality stabilized between generations 200 and 300.
Solution Quality: Across 10 runs, the average best fitness value was 4.2831, with a peak value of 4.3333 achieved in Run 2.
Stability: The algorithm exhibited high robustness, evidenced by a low standard deviation ( $\sigma = 0.0240$ ) in fitness values across different random seeds.
Specific Outcome: The best solution selected 77 resources out of the 217 available candidates, strictly adhering to the time limit while maximizing the pedagogical utility score.

Significance and Future Work
The paper concludes that metaheuristic approaches, specifically the proposed GA with a constraint repair mechanism, are practically viable for real-time decision-support systems in secondary education. The results confirm the algorithm's ability to balance global exploration with local exploitation to find high-quality, feasible solutions for personalized learning pathways.

The authors identify future work in extending this framework to a multi-objective optimization setting. This extension aims to handle students with multiple comorbid diagnoses (e.g., high abilities combined with dyslexia) by formulating distinct objective functions for each diagnosis. The goal would be to resolve potential conflicts between these criteria by identifying Pareto-optimal trade-offs, rather than a single scalar utility score.

Technical Summary: A Genetic Algorithm for Student Academic Resource Allocation

More like this