Simulations and analyses for "Generalizing to generalize: when (and when not) to be compositional in task structure learning"

---

For details explaining the logic of the simulations and experiments, please refer to our paper: bioRxiv

Installation Instructions

This library run on Python 2.7 and unlike most python code, requries compilation with Cython before use. This requires a C compiler (gcc), for which you can find documentation here.

The needed libraries are listed in the file enviornments.yml. If you have conda installed, this file can be used to create a virtual enviornment cython and the other dependencies:
conda env create --file environment.yml

This will create an envionment named "compgen", which you can activate via: source activate compgen

The cython code will need be compiled manually:
python setup.py build_ext --inplace

The entire installation process should take a few (10-15) minutes to complete.

Files List

• For each experiment in the paper, there are two jupyter notebooks, one for simulations (Generative Modeling) and another for human subject behavior (Analysis). The simulations are an example and do not match what was presented in the paper.

• The simulations presented in the paper were not stored for space considerations but can be re-generated. The script batch_run_generative_models.py will regenerate the simulations from seed. To regenerate the plots, the script merge_dataframes.py needs to be run after to prepare the data and then the analysis of the simulations can be run using the notebook Generate figures from batches. All together, this process can take several hours on a modern laptop.

• Exclusion Criteria.ipynb contains the analyses used to exclude individual subjects. There is one notebook for all three

• The folder models contains the code for the computational models (e.g. joint clustering, independent clustering, etc.), as well as the code need to generate experiments for simulations

• opt contains various analysis functions used in the analyesis. This is done to maintain readability in the jupyter notebooks

• data contains deidentified subject data.

• psiTurk contains the code to run the behavioral experiment. These are coded in javascript and run on Amazon's mechanical turk via the psiTurk library [https://psiturk.org].