Multi-Agent Consensus Seeking via
Large Language Models

This file is the source code of our paper "Multi-Agent Consensus Seeking via Large Language Models".

animation.mp4

Prerequisites

Before you begin, ensure you have met the following requirements:

• Python: This project is primarily developed in Python. Make sure you have Python installed on your system. You can download it from python.org.

• Python Libraries: You will need to install several Python libraries to run the code. You can install these libraries using pip, Python's package manager. To install the required libraries, run the following command:

  pip install -r requirements.txt

  The requirements.txt file, included in this repository, lists all the necessary Python libraries along with their versions.

• Docker (Optional): If you plan to deploy the project using Docker, you'll need to have Docker installed on your system.

  1. You can download Docker Desktop from docker.com.

  2. Build the Docker Image:

     To build a Docker image from the included Dockerfile, navigate to the project directory in your terminal and run the following command:

     docker build -t consensus-debate .

     This command will create a Docker image named my-project-image using the Dockerfile in the current directory.

  3. Mounting Your Code:

     You can mount your code into the Docker container by using the -v option when running the container. For example:

     docker run -it -p -v /path/to/your/code:/data consensus-debate /bin/bash

     Replace /path/to/your/code with the absolute path to your project directory. The -v flag maps your local code directory to the /data directory inside the Docker container.

• OpenAI API Keys:

  You will need valid OpenAI API keys to interact with the OpenAI services. Follow these steps to add your API keys to the config.yml file:

  1. Create a config/keys.yml file in the root directory of the project if it doesn't already exist.

  2. Open the config/keys.yml file with a text editor.

  3. Add your API keys in the following format, replacing the placeholder keys with your actual OpenAI API keys:

     api_keys:
       0: "sk-YourFirstAPIKeyHere"
       1: "sk-YourSecondAPIKeyHere"
       2: "sk-YourThirdAPIKeyHere"

     You can add multiple API keys as needed, and they will be accessible by index.

  4. Set the api_base value to the OpenAI API endpoint:

     api_base: 'https://api.openai.com/v1'

  5. Save and close the config.yml file.

By ensuring you have these prerequisites in place, you'll be ready to use the code and run the experiments described in this project.

Running Experiments

1. Create Test Files: In the "test" directory, create one or more Python test files (e.g., my_experiment.py) that define the experiments you want to run. These test files should import and use your Python template as a framework for conducting experiments. Below is an example of what a test file might look like:

   import datetime
   import subprocess
   
   def main(n_agent):
     rounds = 9 # number of rounds in single experiment
     agents = n_agent
     n_stubborn = 0 # number of stubborn agents
     n_suggestible = 0 # number of suggestible agents
     n_exp = 9 # number of experiments
     current_datetime = datetime.datetime.now()
     # Format the date as a string
     formatted_date = current_datetime.strftime("%Y-%m-%d_%H-%M")
     out_file = "./log/scalar_debate/n_agents{}_rounds{}_n_exp{}_{}".format(agents, rounds, n_exp, formatted_date)
     # Build the command line
     cmd = [
       'python', './run.py',
       '--rounds', str(rounds),
       '--out_file', out_file,
       '--agents', str(agents),
       '--n_stubborn', str(n_stubborn),
       '--n_suggestible', str(n_suggestible),
       '--n_exp', str(n_exp),
       # '--not_full_connected' # uncomment this if you want use other topology structures
     ]
   
     # Run the command
     subprocess.run(cmd)
   
   if __name__ == "__main__":
     main(n_agent=3)

   Customize the experiment setup according to your specific needs.

2. Setting the Experiment Type: Before running experiments, make sure to set the appropriate experiment type in the run.py file:

   exp = debate_factory("2d", args, connectivity_matrix=m)

   The first parameter, "2d", specifies the type of experiment. You can use "scalar" for scalar debate or "2d" for vector debate.

3. Run Experiments: You can run the experiments from the command line by executing the test files in the root directory:

   python test/my_experiment.py

   Replace my_experiment.py with the name of the test file you want to run. This will execute your experiment using the Python template and generate results accordingly.

4. Locating Experiment Results: After running experiments using the provided test files, you can find the data files and logs in the "log" directory. The "log" directory is defined in your test files, and it's where your experiment results are stored.

Plotting and Generating HTML

Plotting Data

The code includes functionality for automatic plotting of data when running experiments. However, if you wish to manually plot a specific data file, you can use the following command:

scaler debate:

python -m modules.visual.plot ./log/scalar_debate_temp_0_7/n_agents8_rounds9_n_exp9_2023-10-11_13-44/data.p

vector debate:

python -m modules.visual.plot_2d ./log/vector2d_debate/n_agents3_rounds20_n_exp1_2023-10-27_14-37/trajectory.p

Replace the file path with the path to the specific data file you want to plot. This command will generate plots based on the provided data file.

Generating HTML Reports

The code automatically generates HTML reports for experiments. However, if you want to manually generate an HTML report for a specific experiment or dataset, you can use the following command:

python ./gen_html.py

This command will generate an HTML report based on the data and logs available in the "log" directory.

Please note that for automatically generated HTML reports, the script may take into account the latest experiment data and log files available in the "log" directory. However, running gen_html.py manually allows you to create an HTML report at any time, independently of experiment execution.

Collaborators

• Huaben Chen
• Wenkang Ji

License:

This project is licensed under the MIT License.

Citing

If you find our work useful, please consider citing:

@misc{chen2023multiagent,
  title={Multi-Agent Consensus Seeking via Large Language Models},
  author={Huaben Chen and Wenkang Ji and Lufeng Xu and Shiyu Zhao},
  year={2023},
  eprint={2310.20151},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}