simcat: a Python package to Simulate Multi-agent Cognitive Association Tasks

This is a Python package to perform simulations of multi-agent cognitive association tasks (e.g., the verbal fluency test).

The package makes it eary to instantiate one or more agents, endow them with a semantic memory, defined as a vector space that can be different across agents, and have agents perform association tasks in the form of a verbal fluency tasks either individually of following different interaction structures. At present, the package includes two hard-coded rules:

1. Agents always name the item in their semantic space which has the lowest distance from the prompt (or has the strongest association, depending on options);
2. The simulation stops when all items in the semantic space have been named during a simulation, or when no items in the semantic space can be found, whose distance from the current prompt is lower than (or higher than, depending on options) a set threshold, passed as a parameter to the simulation. These constraints could potentially be relaxed in future releases.

When multiple agents perform the association task together, they can interact following three possible turn-taking rules:

1. Strict turn-taking: agents alternate in naming the next item;
2. Flexible (or collaborative) turn-taking: agents are allowed to step in if the partner has no available associations from the current prompt;
3. Shortest (or competitive) turn-taking: the turn-holder is the agent with the shortest distance between the current prompt and the closest item in their semantic memory. The package is tested for single-agent and two-agent interactions, but it supports interactions with more than two agents. Furthermore, the package makes it possible to endow individual agents with basic forms of working memory, which make it possible to revert to previous prompts if no association from the current prompt is available.

For each simulation, a log of the entire association chain, including a range of metadata (e.g., the distance in semantic space between the prompt and the response item, or the density of the neighborhood in which named words are located), is produced and stored.

🌸 Applications

Applications of this framework include simulation of verbal fluency data, as well as computational studies investigating emergent properties of parameters such as cognitive diversity between agents and turn-taking structures.

A first study using this framework to investigate the effect of cognitive diversity on search behavior and performance in the association task was published in the proceedings of the 2022 Annual Meeting of the Cognitive Science Society (available at: https://escholarship.org/uc/item/58v5d82w):

@inproceedings{rocca2022cognitive,
  title={Cognitive diversity promotes collective creativity: an agent-based simulation},
  author={Rocca, Roberta and Tylén, Kristian},
  booktitle={Proceedings of the Annual Meeting of the Cognitive Science Society},
  volume={44},
  number={44},
  year={2022}
}

A follow-up involving comprehensive investigation of the effect of turn-taking structures and working memory flexibility is available as a preprint at: https://osf.io/preprints/psyarxiv/n3t6j, and currently under review.

@misc{rocca2024diversity,
  title={The effect of diversity and social interaction on cognitive search: An agent-based simulation},
  url={osf.io/preprints/psyarxiv/n3t6j},
  author={Rocca, Roberta and Tylén, Kristian},
  publishr={PsyArXiv},
  year={2024},
}

Materials and documentation concerning this study are available under paper.

⚙️ Installation

We strongly recommend that you install simcat from pip:

pip install simcat

You can also install simcat from source by running:

pip install git+https://github.com/rbroc/simcat.git

or

git clone https://github.com/rbroc/simcat.git
cd simcat
pip3 install -e .

🤖 💬 Running a simulation

Once you have installed simcat can easily run a simulation with the following command:

from simcat import Interaction, Agent

agent_1 = Agent(agent_name="agent_1", matrix_filename="sample/agent_1/matrix.tsv")
agent_2 = Agent(agent_name="agent_2", matrix_filename="sample/agent_2/matrix.tsv")

i = Interaction(
    agents=[agent_1, agent_2],
    save_folder="./sample/outs",
    log_id="my_interaction",
    nr_sim=2,
    map_locations=False,
)

i.run_interaction(seeds=["cat", "dog"], interaction_type="strict", n_back=0)

💾 Simulation outputs

A csv file containing a log of the interaction will be saved, whose first rows will look like this:  

agent	turn	iter	seed	response	prob_a0	prob_a1	threshold	nr_sim	max_exchanges	init_seed	log_id	nr_agents	resp_knnd_5_a0	resp_knnd_5_a0_current	avg_dist_remain_a0	avg_knnd_5_a0	var_knnd_5_a0	resp_neighbors_a0	resp_neighbors_a0_current	resp_knnd_5_a1	resp_knnd_5_a1_current	avg_dist_remain_a1	avg_knnd_5_a1	var_knnd_5_a1	resp_neighbors_a1	resp_neighbors_a1_current
agent_1	0	0	cat	dog	0.00782	0.0113	0.01179	2	240	cat	test_2_agents	2	0.00982	0.0099	0.01274	0.011	0.00071	83	82	0.01053	0.01053	0.01273	0.01099	0.00072	70	69
agent_2	1	0	dog	lion	0.01136	0.00891	0.01179	2	240	cat	test_2_agents	2	0.01066	0.01066	0.01274	0.011	0.00071	50	48	0.00989	0.01023	0.01274	0.011	0.00072	98	96
agent_1	2	0	lion	elephant	0.00938	0.01264	0.01179	2	240	cat	test_2_agents	2	0.00987	0.01007	0.01274	0.01101	0.00071	81	78	0.01148	0.01148	0.01274	0.011	0.00072	10	10
agent_2	3	0	elephant	coral_snake	0.01317	0.01026	0.01179	2	240	cat	test_2_agents	2	0.01019	0.01019	0.01274	0.01101	0.00071	46	46	0.01054	0.01076	0.01274	0.011	0.00072	44	42
agent_1	4	0	coral_snake	sea_snake	0.00836	0.01368	0.01179	2	240	cat	test_2_agents	2	0.01045	0.01053	0.01274	0.01102	0.00071	66	65	0.01056	0.01056	0.01274	0.011	0.00072	27	26
agent_2	5	0	sea_snake	coyote	0.0135	0.00919	0.01179	2	240	cat	test_2_agents	2	0.01071	0.01071	0.01274	0.01102	0.00071	69	67	0.01027	0.01036	0.01274	0.01101	0.00072	40	38
agent_1	6	0	coyote	rabbit	0.00982	0.01201	0.01179	2	240	cat	test_2_agents	2	0.00987	0.01008	0.01275	0.01103	0.00071	91	86	0.01175	0.01175	0.01274	0.01101	0.00072	7	7
agent_2	7	0	rabbit	squid	0.01257	0.01078	0.01179	2	240	cat	test_2_agents	2	0.00955	0.00955	0.01275	0.01103	0.0007	54	53	0.01083	0.01088	0.01274	0.01101	0.00073	54	50
agent_1	8	0	squid	cuttlefish	0.00845	0.01253	0.01179	2	240	cat	test_2_agents	2	0.01011	0.01028	0.01275	0.01104	0.0007	53	50	0.0108	0.0108	0.01274	0.01101	0.00073	54	52
agent_2	9	0	cuttlefish	gull	0.01311	0.01027	0.01179	2	240	cat	test_2_agents	2	0.01164	0.01166	0.01275	0.01104	0.0007	10	9	0.00982	0.00982	0.01274	0.01102	0.00072	83	78

The columns in the output refer to the following:

name	description
agent	name of agent speaking
turn	turn number within the simulation
iter	simulation number
seed	seed at current turn
response	response to seed
prob_a{n}	Euclidean distance between seed and response in the n-th agent's space (Python indexing, therefore the index for the first agent is 0)
threshold	stopping threshold for the simulation
nr_sim	total number of simulations to run
max_exchanges	maximum number of turns within a simulation
init_seed	first seed in the simulation
log_id	id of the simulation, used in the logfile
nr_agents	number of agents
resp_knnd_{k}_a{n}	distance between the response and its k-th nearest neighbor (relative to the original semantic memory of the n-th agent) -- k is passed as a parameter to the Interaction call. This is an index of neighborhood density.
resp_knnd_{k}_a{n}_current	same as above, but computed on the updated space (i.e., accounting only for items that have not been named)
avg_dist_remain_a{n}	average Euclidean distance between each pair of items that have not yet been named (relative to the semantic memory of the n-th agent).
avg_knnd_{k}_a{n}	Mean distance from k-th neighbor, for each item in the n-th agent's semantic space. Only items that have not been named are considered.
var_knnd_{k}_a{n}	Std of the distance from k-th neighbor, for each item in the n-th agent's semantic space. Only items that have not been named are considered.
resp_neighbors_a{n}	number of sub-threshold neighbors of the response item (relative to the semantic memory of the n-th agent)
resp_neighbors_a{n}_current	number of remaining sub-threshold neighbors of the response item (relative to the semantic memory of the n-th agent). Items are dropped from agents' semantic space as they are named.

🛠️ Maintenance and development

Please feel free to contribute to this project. You can contribute to bug fixes 🐛 or implement new functionality 🌱 by opening a PR or opening an issue. You are also very welcome to contribute docs, which are not yet available.

📁 Project structure

simcat
├── LICENSE
├── pyproject.toml
├── README.md
├── img
├── paper      
│   ├── README.md            
│   ├── figures
│   ├── logs
│   ├── metrics
│   ├── models 
│   ├── notebooks 
│   └── scripts
├── sample                  
│   ├── agent_1
│   ├── agent_2
│   └── outs
├── src
│   └── simcat
└── tests

📖 Citation

If you use this work, please cite:

@misc{rocca2024diversity,
  title={The effect of diversity and social interaction on cognitive search: An agent-based simulation},
  url={osf.io/preprints/psyarxiv/n3t6j},
  author={Rocca, Roberta and Tylén, Kristian},
  publishr={PsyArXiv},
  year={2024},
}