Adaptive Spaced Repetition for Vocabulary

ISEF Category: Behavioral and Social Sciences

Subcategory: Cognitive Psychology  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Two students can study the same word list and remember different amounts. The reason may not be effort, it may be timing. Spaced repetition tries to place each review right before you forget, and this project asks whether a small Python agent can do that better than Anki's default SM-2 schedule.

What Is It?

Spaced repetition is a study method that shows you a flashcard, then waits longer before the next review if you remember it, and shorter if you miss it. Anki's default SM-2 schedule is a rule set that does this with fixed math. It works well for many students, but it treats every learner in a similar way.

Your project adds a small reinforcement-learning agent, which is a program that changes its choices after each result. Think of it like a coach that watches your quiz answers and adjusts the next review time. If the agent learns your pattern better than SM-2, your review schedule may get sharper over time.

Why This Is a Good Topic

This is a strong science fair topic because you can measure it with clear numbers, like recall score, retention after a delay, and total reviews needed. You can run it with vocabulary lists, a laptop, and a simple quiz setup, so you do not need a wet lab. It also connects to a real problem, helping students study smarter with less wasted review time.

Research Questions

• How does a personalized interval schedule affect delayed vocabulary recall compared with Anki's default SM-2 schedule?
• What is the effect of a reinforcement-learning agent on the number of words remembered after 4 weeks?
• Does the personalized scheduler reduce the total number of reviews needed to reach a target recall score?
• To what extent does word difficulty change the gap between the two scheduling methods?
• Which kinds of items, new words or review-heavy words, benefit more from personalization?
• How does the personalized schedule change forgetting from the first test to the final test?

Basic Materials

• Laptop or desktop computer with internet access.
• Python installed with pandas, numpy, and matplotlib.
• Anki desktop app with a custom vocabulary deck.
• Spreadsheet software for tracking review logs and quiz scores.
• Matched vocabulary word lists with answer keys and difficulty ratings.
• Google Forms or Microsoft Forms for recall quizzes and response export.

Advanced Materials

• Python with scikit-learn or stable-baselines3 for the agent.
• PsychoPy or Qualtrics for controlled testing.
• Secure data storage for participant records and consent files.
• A larger participant pool, such as classmates or lab volunteers, approved by the lab.
• A statistics package such as R, JASP, or Python statsmodels for mixed-effects analysis.

Software & Tools

• Python: Runs the schedule simulation and analyzes recall data.
• Anki: Provides the SM-2 baseline for direct comparison.
• Jupyter Notebook: Lets you inspect learning curves, errors, and parameter changes.
• Google Sheets: Tracks study sessions, quiz scores, and schedule assignments.
• R: Fits statistical models when you want a second check on the Python results.

Experiment Steps

1. Define the outcome you will score, such as recall accuracy, retention, or reviews per mastered word.
2. Build matched word sets so both schedules start with the same mix of easy and hard items.
3. Decide how the agent will get feedback from each quiz result and how it will compare with SM-2.
4. Plan controls for study time, test timing, and order effects so the schedules stay fair.
5. Predefine the analysis you will use to compare learning curves, forgetting rates, and total review load.
6. Set aside a delayed test or holdout set so you can check whether any gain lasts after the study window.

Common Pitfalls

• Giving the RL agent and SM-2 different starting decks, which makes the final comparison unfair.
• Letting study time vary between conditions, which mixes schedule quality with extra practice.
• Testing only right after a review, which hides whether the schedule helps long-term memory.
• Training the agent on the same data you use to judge it, which inflates the apparent gain.
• Ignoring student-to-student differences, which can hide that one schedule works better for some learners than others.

What Makes This Competitive

A stronger project does more than compare two score averages. It tracks exact review timing, delayed retention, and per-word error patterns, then tests whether personalization helps certain word types more than others. If you also separate training data from evaluation data, your result looks much closer to real research than a simple app comparison.

Project Variations

• Compare personalized scheduling on concrete nouns, abstract nouns, and verb phrases to see whether word type changes the benefit.
• Test the agent against SM-2 on bilingual word pairs instead of single-language vocabulary.
• Replace simple recall with confidence-rated recall and see whether the model learns from uncertainty as well as correctness.

Learn More

• Anki Manual: Read the scheduling and review sections to understand the SM-2 baseline, then find it in Anki's documentation.
• NIH PubMed: Search for review articles on spaced repetition, retrieval practice, and vocabulary learning.
• NIH PubMed Central: Read full-text memory and learning papers for methods and analysis ideas.
• OpenStax Psychology 2e: Review the chapters on memory and learning, available free online.
• MIT OpenCourseWare, Introduction to Psychology: Use the memory units to refresh core learning and forgetting concepts.