AI School Policy Search for Disease Spread

ISEF Category: Computational Biology and Bioinformatics

Subcategory: Computational Epidemiology  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

One school policy can change who gets sick, who falls behind, and who gets left out. That makes this a real optimization problem, not just a yes-or-no debate. You can model those tradeoffs with reinforcement learning, the same idea behind systems that learn from feedback. If you like coding and public health, this project gives you both.

What Is It?

This project asks you to build a computer model of school policy choices and let an algorithm search for better ones. Reinforcement learning means a program tries actions, sees what happens, and learns which choices earn the best score. In your case, the score can combine disease cases, learning loss, and equity, which is a way to measure whether the policy hurts some groups more than others.

The agent-based model is the simulation world. Think of it like a tiny digital city with students, teachers, households, and schools. Each person follows rules for contact, infection, attendance, and testing. You then test different policy plans, like when to close schools, when to require masks, and when to increase testing, and compare how each plan changes the outcome. Calibrating to a real metro area means you tune the model so its population, school structure, or case trends match a real place as closely as your data allow.

Why This Is a Good Topic

This is a strong science fair topic because it has a clear input, a clear output, and a real-world problem behind it. You can change the policy rules, the reward weights, or the fairness metric and measure how the model responds. That makes the project testable, repeatable, and rich enough for deep analysis. You also learn simulation, optimization, and data interpretation, which are valuable skills in computational epidemiology.

Research Questions

• How does changing the weight on learning loss versus case counts alter the policy that reinforcement learning selects?
• What is the effect of adding an equity penalty on the mix of school closure, masking, and testing strategies chosen by the agent?
• Does a policy trained on one metro area still perform well when you test it on a second area with different school size and contact patterns?
• To what extent does increasing testing frequency reduce total cases compared with increasing mask compliance in the model?
• Which reward design produces the most stable policy across repeated simulation runs?
• How does the timing of school closure decisions affect the tradeoff between infection reduction and learning loss?
• What is the effect of using age-stratified contacts instead of uniform contacts on the learned policy?

Basic Materials

• Laptop or desktop computer with at least 8 GB RAM.
• Python installed with NumPy, pandas, matplotlib, and a reinforcement learning library such as Stable Baselines3.
• Spreadsheet software for tracking experiments and results.
• Public data on metro-area population, school enrollment, and case trends from census, state health, or local district sources.
• A notebook for logging model assumptions, reward terms, and policy settings.
• Version control tool such as Git for saving code changes and experiment runs.

Advanced Materials

• Workstation or cloud computing access for repeated simulation runs.
• Python with Mesa or another agent-based modeling framework.
• Python with PyTorch or TensorFlow for custom reinforcement learning experiments.
• Geospatial data for the metro area, including neighborhood or district boundaries.
• Public school attendance, enrollment, or closure data for calibration.
• Statistical analysis tools for sensitivity analysis and uncertainty estimates.

Software & Tools

• Python: Builds the simulation, trains the policy agent, and analyzes results.
• Mesa: Creates the agent-based model of students, schools, and households.
• Stable Baselines3: Provides reinforcement learning algorithms for policy search experiments.
• pandas: Organizes outputs from many simulation runs into clean tables.
• ImageJ: Not useful for this topic, so skip it unless you need image-based data extraction.

Experiment Steps

1. Define the real-world policy question you want your model to answer and decide which outcomes matter most.
2. Choose the population features your model must include, such as school size, contact structure, and attendance rules.
3. Build the reward function so it balances cases, learning loss, and equity in a way you can defend.
4. Calibrate the simulation to public data from one metro area before you start policy search.
5. Compare a reinforcement-learning policy against simple baselines such as fixed rules or threshold rules.
6. Test whether your best policy still works when you change assumptions, data sources, or fairness weights.

Common Pitfalls

• Using a reward function that is too vague, which lets the agent optimize the wrong tradeoff.
• Training on a simulation that does not match real school contacts, which makes the results look precise but weak.
• Ignoring equity until the end, which turns a multi-objective project into a single-outcome model.
• Comparing policies with only one simulation run, which hides how much random outbreak noise affects the results.
• Changing too many model assumptions at once, which makes it hard to tell which policy feature actually caused the outcome.

What Makes This Competitive

A competitive version of this project does more than show that one policy lowers cases. It tests whether the policy still holds under different metro areas, different fairness weights, and different outbreak seeds. Strong projects also compare the learned policy against clear baselines, then explain why the agent picked certain actions. If you add uncertainty analysis and a careful equity measure, your work starts to look like real public health modeling.

Project Variations

• Use a single school district instead of a metro area and compare policy results across elementary, middle, and high school settings.
• Replace the reinforcement-learning agent with a simpler rule-based optimizer and test whether the extra complexity actually improves outcomes.
• Swap the equity metric for absenteeism disparity, then see how the policy changes when you measure fairness in a new way.

Learn More

• NIH PubMed: Search for review articles on agent-based models, school closure policies, and infectious disease decision making.
• CDC data and surveillance pages: Find public outbreak and school health data to inform assumptions and calibration.
• NASA Open Science resources: Learn open modeling and reproducible workflow habits that help with complex simulation projects.
• MIT OpenCourseWare, Introduction to Machine Learning: Review free lecture materials on reinforcement learning concepts and evaluation.
• Mesa documentation: Read the official agent-based modeling framework docs for examples of building and running simulations.
• PLOS Computational Biology: Search for peer-reviewed articles on computational epidemiology, school transmission, and policy modeling.