Inference and search on graph-structured spaces

Table of Contents

• General info
• Data
• Analyses

General info

This repository contains all the code and (anonymized) data necessary for reproducing Wu, Schulz, & Gershman (Computational Brain and Behavior 2021). The following text provides a summary of each file and it's function, with further comments provided throughout the code.

Data

• exp1/ contains a csv of the (anonymized) experiment data experiment1Full.csv and graphs.json defines the graph structures used in experiment 1
• banditTask/ contains a csv of the (anonymized) experiment data networkBandit.csv and network2.json defines the graph structures used in experiment 2

Analyses

• brmsModels/ contains Bayesian mixed effects models. Becuase of the file sizes being potentially larger than the maximum allowed limit in github, all *.brm files are tracked using git-lfs. Please refer to the git-lfs manual to install lfs and pull these files
• modelResults/ contains the individual cross-validated maximum likelihood estimates from Exp 1 (Exp1/) and Exp 2 (networkBandit/). It also contains modelFit.csv and paramEstimates.csv as the compiled dataframes describing the model results in Exp 2 for convenience. In addition, Exp1diffevidence.csv and Exp2diffevidence.csv contain the log loss of each model for computing the protected exceedence probabilities, which is then saved as Exp1PXP.csv and Exp2PXP.csv. Lastly, the model-based analyses of the bonus from from Exp 2 are here for covenience as BanditBonusRoundmodelDF.Rds
• plots/ contains plots for the paper, where the code will save each plot
• rationalModels/ contains data used in the model simulations for Experiment 2. parameters/ holds dataframes for the parameter estimates of each model, where the model simulation (modelSimulations.R) are saved as csv files
• utilities.R contains data preprocessing functions for each experiment and various vector operations that are used across multiple scripts
• statisticalTests.R contains code for performing t-tests and correlations, where the output is formatted for Latex and automatically converted to a set number of significant digits for consistency. Contains code from van Doorn et al., (2018) for computing the Bayes factor for Kendall's rank correlation paper
• exportImportGraph.R contains code by Angelo Antonio Salatino for impoirting and exporting igraph objects as json: original github repo
• models.R contains code defiing models used in both experiments

Experiment 1:

• Exp1Behavior.R contains all the behavioral analyses
• Exp1ModelCV.R contains code for model fitting
• Exp1ModelingResults.R contains code for all model-based analyses

Experiment 2:

• banditExpBehaviorPlots.R contains all the behavioral analyses
• banditModelComparisonCV.R contains code for model fitting
• banditExpBehaviorPlots.R contains code for model-based analyses
• modelSimulations.R contains code for computing the simulated learning curves
• banditBonus.R contains code used for analyzing the bonus round data

PXP

• We provide two different versions of code in both Matlab and Python for computing the protected probability of exceedence (pxp), each producing equivalent results in our analysis.
• The Matlab version is from Sam Gershman, where analysis/bms.m contains the code for computing pxp, while analysis/pxp.m defines the input and output files
• The Python version is from Sichao Yang, where analysis/bms.py contains the code for computing pxp and PXP.ipynb is an interative notebook that loads the inputs and runs it