This is the offical codebase for paper: Smart Agent-Based Modeling: On the Use of Large Language Models in Computer Simulations.

Zengqing, Wu, Run Peng, Xu Han, Shuyuan Zheng, Yixin Zhang, Chuan Xiao. 2023.

We recommend using anaconda to build a virtual environment to test our codes.

git clone https://github.com/Roihn/SABM.git
cd sabm
conda create -n sabm python=3.10
conda activate sabm
pip install -r requirements.txt

Please make sure you have the valid openai api key for GPT4. Otherwise, you may need to apply for one before you test our codes.

Two options are prepared for you to easily apply your api key to our codebase. You can choose either one of them, or both.

1. Setup environment variable

   Run the following command in your terminal.

   export OPENAI_API_KEY=<your-api-key>

   You may need to run it everytime you open a new terminal before you test our codes.

2. Local API storage

   This is a one-step setup. You only need to paste your api key in the apikey.token file. Then, you can test our codes without any extra setup.

   *Please do not commit your api key to github, or share it with anyone online.

To reproduce the results in the paper, please run the following command.

python main_number_guessing.py

You can specify additional parameters to reproduce the results of the paper.

The <persona> can be "default" (no persona), "aggressive", or "conservative".

The <advanced> option is used to specify advanced agent modeling options, including "domain_knowledge", "learning", "reasoning", "planning", and "hint" (conversation) discussed in the paper.

The <set_guess_number> is used to guess a fixed number.

The <interpretation_guess> is used for model interpretation.

python main_number_guessing.py --model_version "gpt-4-0613" --persona "conservative" --advanced "domain_knowledge" --set_guess_number --interpretation_guess

To reproduce the results in the paper, please run the following command.

python main_emergency_evacuation.py --task <task_id>

The <task_id> can be one of the following options: [1, 2, 3, 4]. Each task id corresponds to a specific case study in the paper.

Additionally, if you would like to specify a seed, the number of humans, and add obstacles in the environment, you may want to run the following command.

python main_emergency_evacuation.py --task <task_id> --seed <seed> --num_humans <num_humans> --need_obstacle

To reproduce the results in the paper, please run the following command.

python main_plea_bargain.py

We provide a GUI to set the parameters of the run.

python main_plea_bargain.py --gui

If you choose to use command instead of the GUI to set the parameters for simulation, you may want to run the following command.

The <persona> can be "persona" or "nopersona", indicating whether or not the persona is used in the performance of plea bargaining.

The <no_fewshot> option indicates not to provide a few-shot context to the agent in the plea bargain.

The <task> can be one of the following options: [1, 2, 3]. Each task id corresponds to a specific case study in the paper.

python main_plea_bargain.py --model_version "gpt-4-0314" --tcu_test --persona "nopersona" --no_fewshot --output_max_tokens 100 --num_agents 1 --task 1

To reproduce the results in the paper, please run the following command.

python main_firm_pricing_competition.py

We provide a GUI to set the parameters of the run.

python main_firm_pricing_competition.py --gui

If you choose to use command instead of the GUI to set the parameters for simulation, you may want to run the following command. The <persona_firm> parameter accepts one of the following options: [0, 1, 2]. Here, 0 indicates no persona, 1 represents an active persona, and 2 denotes an aggressive persona.

python main_firm_pricing_competition.py --model_version "gpt-4-0314" --rounds 1000 --output_max_tokens 100 --breakpoint_rounds 20 --persona_firm1 1 --persona_firm2 1 --set_initial_price --cost 2 2 --parameter_a 14 --parameter_d 0.00333333333333 --parameter_beta 0.00666666666666 --initial_price 2 2 --load_data_location "Record-231112-1955-gpt-4-0314" --strategy --has_conversation

If you find our work useful, please give us credit by citing:

@misc{wu2023smart,
      title={Smart Agent-Based Modeling: On the Use of Large Language Models in Computer Simulations}, 
      author={Zengqing Wu and Run Peng and Xu Han and Shuyuan Zheng and Yixin Zhang and Chuan Xiao},
      year={2023},
      eprint={2311.06330},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}