Sequential rule learning

Abstract

This research study explores the human capacity for generalizing abstract sequential rules using a novel learning paradigm. Participants were presented with a task where they had to predict the outcome of 2-step transformations. The study compares the effects of a simple curriculum and a complex curriculum on learning performance and generalization. The results indicate that the simple curriculum leads to faster initial learning and higher performance, but neither curriculum shows a transfer effect on average. Model-based analysis identified distinct response patterns among participants, and when categorizing them based on intermediate levels of rule comprehension, a difference in transfer ability was observed between the training curricula. Specifically, participants who learned the ground-truth rule in the complex curriculum exhibited superior transfer compared to those in the simple condition. We ran six different versions of this experiment on prolific, the final one consisting of two sessions.

This repository contains the preprocessed data and the scripts that recreate the figures in the report.

A detailed report of all experiments conducted and results can be found here

Folder Structure

The project is organized as follows:

• data: Contains the preprocessed data used in the research.
  • all_data.json: contains the preprocessed data
  • perppt_best_modelfit.json: contains a table with model fits for each pp
  • all_data_tagged.csv: contains a table with manually assigned tag based on ppt debrief
• scripts: Contains the main code for data preprocessing, model-fitting, and visualization.
  • seqlearn_preprocessing.py: Cleaning of data
  • seqlearn_datacollection_bookkeeping.py: Outputs stats on experiment 7 data collection
  • seqlearn_errors_datadriven.py: Exploratory clustering analysis based on error patterns
  • seqlearn_errors_modelbased.py: Model fitting analysis
  • seqlearn_plotting.py: Generates visualizations based on the processed data.
  • seqlearn_tag_data_manual.py: Script used to tag each debrief response.
• results: Contains the figures and outputs generated by the code.
• docs: Contains documentation and supplementary materials for the research project.
  • data_description.md: Provides detailed information about the dataset, its variables, and their meanings.
  • methodology.md: Explains the research methodology, including the statistical models and algorithms used.
• LICENSE: Specifies the license under which the code and data are released.
• README.md: The main README file containing project information.

Requirements

• matplotlib==3.5.1
• numpy==1.21.2
• pandas==1.3.4
• scikit_learn==1.0.2
• scipy==1.7.3
• scripts==2.0
• seaborn==0.11.2