Doing Experiments and Revising Rules with Natural Language and Probabilistic Reasoning [NeurIPS 2024]

Paper on Arxiv -- Accepted at NeurIPS 2024

Installation

Create a conda environment and install the requirements and openai-hf-interface submodule

conda create -n de_rr python=3.8
conda activate de_rr
pip install -r requirements.txt
cd openai-hf-interface
pip install -e .

To run our algorithms, you need to provide OpenAI API key because our algorithms use gpt-4. We use openai-hf-interface, a prompting package that handles sending requests to LLM APIs, in this repo, so you can provide your OpenAI API key by creating secrets.json file inside openai-hf-interface directory. The json file should have:

{
    "openai_api_key": "put-your-key-here",
}

Please see openai-hf-interface for more information.

Running

Note: please use the acre branch to run the ActiveACRE experiments and use the main branch to run the Zendo experiments.

Call the below command to run an experiment:

python run.py agent/proposal=particle_filter # To run online, fuzzy method on Zendo
python run.py dataset=acre agent=acre agent/proposal=particle_filter # To run online, fuzzy method on ActiveACRE