This file was generated on 2025-08-20 by Sabrina Walch

A GENERAL INFORMATION

1.  Title of the dataset:"A framework for the emergence and analysis of language in social learning agents"

2.  Brief description of the research project and its aims: This research project investigates the emergence of language in a multi-agent reinforcement learning system.  It explores how a teacher agent can communicate task-relevant information to a student agent to improve task performance and generalization. The study analyzes the structure of the emergent language and its relationship to internal representations and task performance. The aim is to understand the role of language as a shared representation among agents and its implications for generalization capabilities.

3.  Author Information

    A.  *Investigator Contact Information*
        Name: Tobias J. Wieczorek
        Institution: Department of Computer Science, Technical University Darmstadt
        Address: Darmstadt, Germany (See paper for complete address details)
        Email: See paper for contact information

    B.  *Project Supervisor (Principal Investigator) Contact Information*
        Name: Carlos Wert-Carvajal & Maximilian F. Eggl
        Institution: Institute of Experimental Epileptology and Cognition Research, University of Bonn Medical Center
        Address: Bonn, Germany (See paper for complete address details)
        Email: See paper for contact information

    C.  *In case of questions related to this dataset, please contact:*
        Name: Carlos Wert-Carvajal or Maximilian F. Eggl
        Institution: Institute of Experimental Epileptology and Cognition Research, University of Bonn Medical Center
        Address: Bonn, Germany (See paper for complete address details)
        Email: See paper for contact information

4.  Date of data collection: 2022-05-30 To 2024-08-16 

5.  Information about funding sources that supported the collection of the data:
    Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)-Project-ID 227953431-SFB 1089

6.  Language of the dataset:
    Python (code), Numerical data (Q-matrices, etc.)

7.  Geographic location of data collection:
    Darmstadt, Hessen, Germany & Bonn, Nordrhein-Westfalen, Germany

# B DATA & FILE OVERVIEW

1.  File List:

    *   `code/`: Folder containing the Python code for training agents, autoencoder, and analysis.
    *   `data/`: Folder containing the Q-matrices, encoded messages, and student performance data.
    *   `models/`: Folder containing trained models for teacher, autoencoder, and student agents.
    *   `README.md`: This file.

2.  Are there multiple versions of the dataset? no

3.  Relationship between files:
    The code in `code/` generates the data in `data/` and uses the models in `models/`. The data in `data/` is used to analyze the emergent language and evaluate the student performance.

4.  Additional related data collected that was not included in the current data package:
    None specified in the article.

# C SHARING/ACCESS INFORMATION

1.  Was data derived from another source? no

2.  Licenses/restrictions placed on the data:
    Specify the license, e.g., MIT License (refer to github.com/megg123/multi-agent_language for details)

3.  Links to publications that cite or use the data:
    https://doi.org/10.1038/s41467-024-51887-5

4.  Links to other publicly accessible locations of the data:
    github.com/megg123/multi-agent_language
    zenodo.org/doi/10.5281/zenodo.7885526

5.  Links/relationships to ancillary datasets:
    None specified in the article.

# D METHODOLOGICAL INFORMATION

1.  Description of methods used for collection/generation of data:

    *   Teacher Agent Training: Deep Q-learning with a multi-layer perceptron to approximate the Q-function.
    *   Sparse Autoencoder Training: Minimizing a loss function consisting of reconstruction loss, sparsity regularization, and (optionally) a term that maximizes student task success.
    *   Student Agent Training: Learning to interpret messages and solve the navigation task using reinforcement learning.
    *   Analysis: Principal Component Analysis (PCA), topographic similarity analysis, and entropy analysis were used to analyze the emergent language.

2.  Methods for processing the data:
    Q-matrices generated by the teacher agent are compressed by the sparse autoencoder to generate messages. The student agent uses these messages to learn policies for solving tasks. Data collected during this process, such as rewards and actions, is used to evaluate performance. Analyzed using PCA to derive insight from it

3.  Instrument- and/or software-specific information needed to interpret the data:

    *   Python 3.x
    *   TensorFlow/PyTorch [Specific version numbers would be ideal]
    *   NumPy
    *   Scikit-learn

4.  People involved in sample collection, processing, analysis and/or submission:
    Tobias J. Wieczorek, Carlos Wert-Carvajal, Maximilian F. Eggl, Tatjana Tchumatchenko

5.  Describe any quality-assurance procedures performed on the data:
    The results were statistically validated with t-tests and variance analysis.

6.  Standards and calibration information:
    Not applicable

7.  Environmental/experimental conditions:
    Simulated grid-world environment.

# E DATA-SPECIFIC INFORMATION FOR: [FILENAME]

    Replace this with the filename for each dataset, folder, or file

    Example: E DATA-SPECIFIC INFORMATION FOR: `data/Q_matrices.csv`

    1.  Variable list including full names and definitions of column headings for tabular data:

        *   `state_x`: X-coordinate of the agent's state.
        *   `state_y`: Y-coordinate of the agent's state.
        *   `action_up`: Q-value for moving up in that state.
        *   `action_down`: Q-value for moving down in that state.
        *   `action_left`: Q-value for moving left in that state.
        *   `action_right`: Q-value for moving right in that state.

    2.  Units of measurement used:
        Q-values are dimensionless, representing expected future reward.

    3.  Missing data codes/symbols:
        `NA`: Indicates that a state is inaccessible (e.g., a wall).

    4.  Specialized formats or other abbreviations used:
        None specified.