This repository contains the source code used for experiments in the following paper:

Cranefield, Stephen and Dhiman, Ashish. Identifying Norms from Observation Using MCMC Sampling, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, International Joint Conferences on Artificial Intelligence, 118-124, 2021. https://doi.org/10.24963/ijcai.2021/17 (BibTeX)

The supplementary material for the paper can be found at Learning_Norms_with_MCMC_supp_material.pdf

The project covers the steps as given in the schematic below:

To clone and run this application, you'll need to follow the below-mentioned steps:

# Clone this repository
$ git clone https://github.com/ashish1610dhiman/learning_norms_with_mcmc_from_pcfg_IJCAI21

# Go into the repository
$ cd learning_norms_with_mcmc_from_pcfg_IJCAI21

# Ensure the repository is in the correct branch for the IJCAI experiments
$ git checkout non_compliance

# Install dependencies using pip
$ pip install -r requirements.txt

# Or install dependencies in a virtual environment, e.g. for conda:
$ conda create --name <env_name> --file requirements.txt

├── LICENSE
|
├── README.md               <- The top-level README.
|
├── *_supp_material.pdf     <- Supplemetary material for paper published in IJCAI-21.
|
├── data_nc/*               <- Folder with dvc files for various experiments with p_nn > 0
|
├── data/*                  <- Folder with dvc files for various experiments with p_nn = 0
|
├── src/
│   ├── mcmc_norm_*/        <- Code files for grammar/Metropolis Hastings Algorithm/convergence
|   |                          and preciscion-recall
│   └── *.py                <- Small Helper files
│
├── scripts/                <- Scripts used for variouis instances of the process depicted in 
|   |                          schematic above.
│   └── nc_experiments.py   <- Binding script used to run various parts of experiment
|
├── notebooks/              <- Jupyter notebooks. Naming convention is a number (for ordering),
│                              the creator's initials, and a short `_` delimited description. The notebooks with tag 1.5 mark 
|                              the files used for experiment shown in paper.
|
├── params_nc.yaml          <- yaml file detailing parameters for experiments used
│
└── requirements.txt        <- The requirements file for reproducing the analysis environment

The project uses dvc for tracking data. There are two data folders in the repository:

1. data/: houses dvc files for data produced with older experiments
2. data_nc/: dvc files for data produced with IJCAI submission experiments. The dvc files follow naming as per the iteration of experiment, or exp_nc{i}: for i ∈ {1,2,... , 5}

exp_nc5.dvc is the file corresponding to experiment presented in paper. Earlier experiments were for testing code changes.

Use the following commands to download the data corresponding to a dvc file:

# fetch the data
$ dvc fetch exp_nc5.dvc

# checkout the branch
$ dvc checkout

As outlined in the About The Project above, MCMC Norm learning pipleine involves the following steps:

1. Initialise a Environment and task
2. Generate obervations
3. Generate MCMC chains
4. Analyse MCMC chains
5. Run Convergence tests on MCMC chains
6. Get Top Norms from chains

The above steps are outlined in the binding script nc_experiments.py.

Jupyter notebook is then used as a wrapper over nc_experiments.py, to run different experiment iterations. The naming scheme of notebooks is mentioned above in Project Organsisation.

1.5_nc_exp5.ipynb is the notebook used to run experiment presented in paper

1. Stephen Cranefield
   Department of Information Science, University of Otago (Professor)
   Google Scholar

2. Ashish Dhiman
   Department of Aerospace Engineering, Indian Institute of Technology, Kharagpur (B-Tech, 2019)
   Connect on LinkedIn