tomov/autoneuro-synthetic

Synthetic data generation for the automated neuroscientist

Agents and environments for generating synthetic data for the automated neuroscientist.

Setup

conda create -n synthetic python=3.9
conda activate synthetic
pip install -r requirements.txt

Generate data

Classical conditioning

python -m experiments.gen_classical

Need to edit file to change environments and/or agents.

Data

Classical conditioning

Directory structure: data/classical/<behavioral phenomenon>/

Example:

data/
  classical/
    overshadowing/  
    recovery/  
    second_order_LI/  
    ...

Each has subdirectories:

• For input variables: input/<variable name>.csv
• For output variables: output/<model name>/<variable name>.csv

Each file corresponds to a different variable. Each row corresponds to a (sequential) data point.

Concatenating all the files (horizontally) for a given behavioral phenomenon (e.g., data/classical/overshadowing/) produces a dataset for the automated neuroscientist.

Example for Rescorla-Wagner model:

input/
  states.csv
  rewards.csv
output/
  RW/
    rpes.csv
    values.csv

In this case, we're trying to recover the function y = f(x), where:

• x = [state, reward]
• y = [rpe, value]

Instrumental conditioning

Directory structure: data/instrumental/<env>/<agent and params/

To generate, edit the main in experiments/gen_instrumental.py (TODO make configurable). NOTE: will overwrite data for same settings.

Run from root:

python -m experiments.gen_instrumental