MLDR: A Machine Learning-driven Radio Interface

MLDR: A Machine Learning-driven Radio Interface Version 1 Description 1. Introduction This document contains the data management plan of the MLDR: A Machine Learningdriven Radio (MLDR) Interface project. It reports the data types the project will be dealing with, how they will be stored and released / disseminated. Special attention is placed on the open access publication policy the project will follow. The structure of this report is based on the Data Management Plan (DMP) submitted by a project partner (AGH University) to its national funding agency (i.e., National Science Centre, NCN). MLDR is a project from the CHIST-ERA 2022 Call — Wireless Artificial Intelligence topic. It started in February 2024, and will finish in January 2027. More details about the project can be found in the project web page: https://www.upf.edu/web/mldr.eu As indicated earlier, Publications are the most discussed and understood data. MLDR adheres to Open Access publication as a matter of principle, and compliance with Horizon Europe rules for Open Access to scientific publications needs to be ensured. For each publication we will choose the most suitable approach (either “green” OA or “gold” OA) for each publication concerned. LICENSE:CC-BY-4.0 DOI: - 04/12/2025 The e-repository, which is the institutional repository of UPF, the coordinator of the MLDR project, is among the top 20 providers of publications and projects in H2020 and ERC: it meets all the requirements established by the European Union within the framework of Open Access publishing, and will be used for UPF co-authored publications. Regarding Datasets. The project will follow the principle of "as open as possible, as closed as necessary" for its data, metadata, and software under the FAIR principles to make data and software findable, accessible, interoperable and reusable. The project open metadata will be licensed under CC or equivalent. This DMP is released by April 2024, and will be updated—if required—by the end of the project. 2. Data Management To make research data findable, accessible, interoperable and re-usable (FAIR), the DMP includes information on: a) what data will be collected, processed and/or generated; Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 b) the handling of any kind of data accessible by/from the project during and after the end of the project; c) which methodology and standards will be applied; d) whether any resulting data will be shared/made open access; and e) how data will be curated and preserved (including after the end of the project). 2.1 Data description and collection or re-use of existing data What data (for example the types, formats, and volumes) will be collected or produced? MLDR will generate the following types of data: Temporal data describing the evolution of a certain protocol / functionality / system; Use-case descriptions and KPIs; Specific data examples include: channel occupancy, packet error rate, received signal, CSI angles, etc. ML algorithms. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 The MLDR evaluation framework (codes and models) will be published as software, including details to reproduce the same experiments done during the project execution. A detailed implementation of the use-cases, and datasets obtained will be also made publicly available. Datasets will be published describing how they were obtained (experimental setup and environmental conditions). In case they were obtained using the MLDR evaluation framework, we will explicitly provide the necessary scripts to generate them. An example dataset made available by the consortium is the WACA FCB dataset containing spectrum measurements. Regarding ML algorithms, they will be published as software modules and made available to other researchers so that they can reproduce the results of the project as well as apply them to other problems. The data is classified into different types, namely, publications, data sets and tools. We will also have all the documentation generated by the project. In detail, the project will create new data in the form of: Scientific publications. Tools and libraries, i.e., simulation modules, implementation of algorithms, data processing routines, etc. Source code, from abovementioned tools and libraries. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 Experimental (raw) results - the output of running experiments, e.g., as a CSV file with numerical values. Measurements from real systems, e.g., RSSI values. Analyzed results - the output of using data analysis tools (e.g., custom Python scripts) to analyze the raw, experimental results. Project documentation: slides and reports. Any newly created data will be subject to validity tests in which the outcome is expected. For example, a newly developed simulation scenario (source code) will be run in settings where both the raw and analyzed results can be compared with results from the literature. All experiments will be run multiple times (if it proceeds) and the results will be subject to a statistical analysis to ensure their validity. MLDR does not require data from external sources. However, we are open to using external datasets if they can contribute to achieve the goals of the project, e.g., to pretrain a certain AI/ML module. The project will reuse publicly available open-source applications (e.g., existing simulators), libraries (e.g., matplotlib), and implementations (e.g., of machine learning algorithms). How data will be stored? Which data formats will be used? Data will be stored on machines—either physical or virtual—dedicated to support the project. When required, the machine will be running a server of a distributed version control system (e.g., git) which will encompass all the data produced in the project. This Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 system will ensure that revisions of all files are kept and that the contribution of each project member can be documented. Regarding the storage and curation of the research data generated and/or collected during the project, we plan the following: Repositories: Github, Zenodo, IEEE Dataport, arXiv, TechRxiv etc. Preservation: All listed repositories guarantee data preservation. For project documentation, and sharing data between partners, Google Drive will be used. It provides all access and security requirements, as well as data management services (back-up, multiple copies, version control, etc.) We will put a strong emphasis in the project on using text-based formats, which ensure full readability: source code: .cc, .py results: .csv figures: .svg articles: .tex, .bib Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 These formats do not require much space, usually < 1 GB. More space is required by the open-source applications, e.g., ns-3 requires about 2 GB. Nonetheless, the project will not consume much storage capacity. 2.2 Documentation and data quality What metadata and documentation (for example methodology or data collection and way of organising data) will accompany data? The project will require storing meta-data about the experimental results, which will provide the following information: version of code used to run experiment, experiment settings. We will store this information using text-based files and (if possible) using dedicated opensource software, e.g., sacred, to keep track of all the information required to reproduce the results. Software such as sacred uses a dedicated database for storing results, with links to the results. For the text-based approach, we will use the following folder structure: %data-%id-%name-%config\ where the variables are, respectively, date of experiment, unique experiment id (consecutive number), experiment name, unique configuration name. Each folder will contain a human-readable README file. Furthermore, we plan to adopt an open source and text-based meta-data description specification such as Data Packages. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 What data quality control measures will be used? The tools used to generate data in the project are mostly based on pseudo-random number generators (which means they are fully reproducible) and on deterministic data analysis. Therefore, the acquiring, processing, and analysis of data does not impact the quality of the data. Also, the obtained data does not require cleaning. Nonetheless, the data can be erroneous due to errors in implementation (of simulation scenarios or ML algorithms). Software testing methods and cross-validation with other tools (where available) will be used to eliminate such errors. 2.3 Storage and backup during the research process How will data and metadata be stored and backed up during the research process? Google Drive will be used to store and backup data. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 When applicable, data and meta-data will be stored on dedicated machines with a version control system. The machine will be regularly and automatically backed up using procedures provided by the hypervisor. It will allow rescuing data in case of an incident. How will data security and protection of sensitive data be taken care of during the research? Google Drive provides the required access control and security mechanisms to protect the generated data. In case the project's machines are used, they will only be accessible through secured protocols. In any of the two cases, access will be limited only to project personnel. The project does not encompass any sensitive data. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 1. Basic Information Title: MLDR: A Machine Learning-driven Radio Interface Description: 1. Introduction This document contains the data management plan of the MLDR: A Machine Learningdriven Radio (MLDR) Interface project. It reports the data types the project will be dealing with, how they will be stored and released / disseminated. Special attention is placed on the open access publication policy the project will follow. The structure of this report is based on the Data Management Plan (DMP) submitted by a project partner (AGH University) to its national funding agency (i.e., National Science Centre, NCN). MLDR is a project from the CHIST-ERA 2022 Call — Wireless Artificial Intelligence topic. It started in February 2024, and will finish in January 2027. More details about the project can be found in the project web page: https://www.upf.edu/web/mldr.eu As indicated earlier, Publications are the most discussed and understood data. MLDR adheres to Open Access publication as a matter of principle, and compliance with Horizon Europe rules for Open Access to scientific publications needs to be ensured. For each publication we will choose the most suitable approach (either “green” OA or “gold” OA) for each publication concerned. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 The e-repository, which is the institutional repository of UPF, the coordinator of the MLDR project, is among the top 20 providers of publications and projects in H2020 and ERC: it meets all the requirements established by the European Union within the framework of Open Access publishing, and will be used for UPF co-authored publications. Regarding Datasets. The project will follow the principle of "as open as possible, as closed as necessary" for its data, metadata, and software under the FAIR principles to make data and software findable, accessible, interoperable and reusable. The project open metadata will be licensed under CC or equivalent. This DMP is released by April 2024, and will be updated—if required—by the end of the project. 2. Data Management To make research data findable, accessible, interoperable and re-usable (FAIR), the DMP includes information on: a) what data will be collected, processed and/or generated; b) the handling of any kind of data accessible by/from the project during and after the end of the project; c) which methodology and standards will be applied; Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 d) whether any resulting data will be shared/made open access; and e) how data will be curated and preserved (including after the end of the project). 2.1 Data description and collection or re-use of existing data What data (for example the types, formats, and volumes) will be collected or produced? MLDR will generate the following types of data: Temporal data describing the evolution of a certain protocol / functionality / system; Use-case descriptions and KPIs; Specific data examples include: channel occupancy, packet error rate, received signal, CSI angles, etc. ML algorithms. The MLDR evaluation framework (codes and models) will be published as software, including details to reproduce the same experiments done during the project execution. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 A detailed implementation of the use-cases, and datasets obtained will be also made publicly available. Datasets will be published describing how they were obtained (experimental setup and environmental conditions). In case they were obtained using the MLDR evaluation framework, we will explicitly provide the necessary scripts to generate them. An example dataset made available by the consortium is the WACA FCB dataset containing spectrum measurements. Regarding ML algorithms, they will be published as software modules and made available to other researchers so that they can reproduce the results of the project as well as apply them to other problems. The data is classified into different types, namely, publications, data sets and tools. We will also have all the documentation generated by the project. In detail, the project will create new data in the form of: Scientific publications. Tools and libraries, i.e., simulation modules, implementation of algorithms, data processing routines, etc. Source code, from abovementioned tools and libraries. Experimental (raw) results - the output of running experiments, e.g., as a CSV file with numerical values. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 Measurements from real systems, e.g., RSSI values. Analyzed results - the output of using data analysis tools (e.g., custom Python scripts) to analyze the raw, experimental results. Project documentation: slides and reports. Any newly created data will be subject to validity tests in which the outcome is expected. For example, a newly developed simulation scenario (source code) will be run in settings where both the raw and analyzed results can be compared with results from the literature. All experiments will be run multiple times (if it proceeds) and the results will be subject to a statistical analysis to ensure their validity. MLDR does not require data from external sources. However, we are open to using external datasets if they can contribute to achieve the goals of the project, e.g., to pretrain a certain AI/ML module. The project will reuse publicly available open-source applications (e.g., existing simulators), libraries (e.g., matplotlib), and implementations (e.g., of machine learning algorithms). How data will be stored? Which data formats will be used? Data will be stored on machines—either physical or virtual—dedicated to support the project. When required, the machine will be running a server of a distributed version control system (e.g., git) which will encompass all the data produced in the project. This system will ensure that revisions of all files are kept and that the contribution of each project member can be documented. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 Regarding the storage and curation of the research data generated and/or collected during the project, we plan the following: Repositories: Github, Zenodo, IEEE Dataport, arXiv, TechRxiv etc. Preservation: All listed repositories guarantee data preservation. For project documentation, and sharing data between partners, Google Drive will be used. It provides all access and security requirements, as well as data management services (back-up, multiple copies, version control, etc.) We will put a strong emphasis in the project on using text-based formats, which ensure full readability: source code: .cc, .py results: .csv figures: .svg articles: .tex, .bib These formats do not require much space, usually < 1 GB. More space is required by the open-source applications, e.g., ns-3 requires about 2 GB. Nonetheless, the project will not consume much storage capacity. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 2.2 Documentation and data quality What metadata and documentation (for example methodology or data collection and way of organising data) will accompany data? The project will require storing meta-data about the experimental results, which will provide the following information: version of code used to run experiment, experiment settings. We will store this information using text-based files and (if possible) using dedicated opensource software, e.g., sacred, to keep track of all the information required to reproduce the results. Software such as sacred uses a dedicated database for storing results, with links to the results. For the text-based approach, we will use the following folder structure: %data-%id-%name-%config\ where the variables are, respectively, date of experiment, unique experiment id (consecutive number), experiment name, unique configuration name. Each folder will contain a human-readable README file. Furthermore, we plan to adopt an open source and text-based meta-data description specification such as Data Packages. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 What data quality control measures will be used? The tools used to generate data in the project are mostly based on pseudo-random number generators (which means they are fully reproducible) and on deterministic data analysis. Therefore, the acquiring, processing, and analysis of data does not impact the quality of the data. Also, the obtained data does not require cleaning. Nonetheless, the data can be erroneous due to errors in implementation (of simulation scenarios or ML algorithms). Software testing methods and cross-validation with other tools (where available) will be used to eliminate such errors. 2.3 Storage and backup during the research process How will data and metadata be stored and backed up during the research process? Google Drive will be used to store and backup data. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 When applicable, data and meta-data will be stored on dedicated machines with a version control system. The machine will be regularly and automatically backed up using procedures provided by the hypervisor. It will allow rescuing data in case of an incident. How will data security and protection of sensitive data be taken care of during the research? Google Drive provides the required access control and security mechanisms to protect the generated data. In case the project's machines are used, they will only be accessible through secured protocols. In any of the two cases, access will be limited only to project personnel. The project does not encompass any sensitive data. 2.4 Legal requirements, codes of conduct Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 If personal data are processed, how will compliance with legislation on personal data and on data security be ensured? No personal data is expected to be used. In the eventual case the use of personal data is required, it will be handled with appropriate ethical procedures. If needed for the implementation of (specific elements of) the MLDR project, the academic partners will acquire approval from their respective ethical committees. How will other legal issues, such as intelectual property rights and ownership, be managed? What legislation is applicable? We will follow the Consortium Agreement signed by MLDR partners. 2.5. Data sharing and long-term preservation Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 Simulation (e.g., climate modeling data) 1.1.2 Give details on the data format The dataset is provided entirely in open, non-proprietary formats that ensure longterm usability and interoperability. The core simulation logs are stored as CSV files, which contain per-learning-AP statistics (including agent actions, delays, and rewards) as well as per-station reception statistics (including packet delays and goodput), all aligned to a common temporal reference. Runtime information from WandB, including execution metadata, step logs, and timestamps, is also provided as CSV files. Structured experiment summaries generated at the end of each simulation are stored in JSON files, capturing hierarchical metadata and aggregated metrics in a humanand machine-readable format. To preserve the complete WandB history of each simulation efficiently, the dataset includes Pickle (PKL) files. In summary, the dataset includes the following formats: •CSV (learning APs: actions, delays, and rewards; stations: goodput and delays) •JSON (simulation summaries) •PKL (full WandB histories) •CSV (WandB runtime, step, and timestamp metadata) 1.1.3 Justify the use of certain formats widely supported format across disciplines CSV files are widely used plain-text formats for tabular data, making them suitable for logging large volumes of simulation statistics in a simple and tool-agnostic Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 manner. JSON files provide a structured, readable, and extensible format, ideal for storing aggregated metrics and metadata. PKL files allow efficient serialization of Python objects (such as dataframes) preserving complex data structures and enabling fast loading. Together, these open, non-proprietary formats ensure the dataset remains accessible, interoperable, and fully reproducible. 1.1.4 Give details on the volumes GB (gigabyte) The dataset has a total size of approximately 5.1 GB, including all scenarios, algorithms, and seeds. Individual CSV files typically fall within the hundreds of kilobytes. JSON summary files are generally on the order of tens of kilobytes. PKL files, in contrast, are usually on the order of tens of megabytes. 1.2 How will new data be collected or produced? 1.2.1 Explain which methodologies or software will be used if new data are collected or produced or if third party data are used The dataset was produced using WiPySim, an event-driven IEEE 802.11 simulator implemented in Python. Simulation outputs and logs were automatically collected and logged using WandB. 2 Documentation And Data Quality 2.1 Documentation And Data Quality 2.1.1 What metadata and documentation will accompany the data? 2.1.1.3 Indicate how the data will be organised during the project Data is systematically organized in a hierarchical folder structure. Each scenario has its own directory, with subdirectories for each algorithm and architecture combination. Individual seeds for repeated simulation runs are stored in separate folders, each containing per-node directories. Naming conventions are consistent, using identifiers such as node_X for learning APs and node_10X for associated STAs (e.g., node_1 paired with node_101), and seed_X for random seeds. This structure, combined with consistent and descriptive filenames, ensures that all data are easily navigable, well organized, and ready for reuse. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 3 Reused Data 4 Storage And Backup During The Research Process 4.1 Storage And Backup During The Research Process 4.1.1 How will data security and protection of sensitive data be taken care of during the research? 4.1.1.1 Explain how the data will be recovered in the event of an incident and describe the main risks and how these will be managed The dataset contains no personal or sensitive information, and its recovery relies on externally hosted services. The complete and fully structured dataset is permanently archived on Zenodo, which assigns a DOI and ensures longterm preservation, versioning, and recoverability. In addition, WandB maintains a private, non-publicly accessible record of each simulation run, storing raw experiment metadata and logs that can be retrieved or reexported if needed. Zenodo therefore provides the openly available, curated dataset for long-term access, while WandB serves only as an internal auxiliary source. 5 Legal And Ethical Requirements, Codes Of Conduct 5.1 Legal And Ethical Requirements, Codes Of Conduct 5.1.1 Personal data 5.1.1.1 Are there any personal data to be formulated? No 5.1.2 How will other legal issues, such as intellectual property rights and ownership, be managed? What legislation is applicable? 5.1.2.1 Data ownership and accessibility 5.1.2.1.1 Who will be the owner(s) of the data? Universitat Pompeu Fabra • Boris Bellalta (orcid: 0000-0002-7826-3703) • Miguel Casasnovas (0009-0000-4787-730X) • Francesc Wilhelmi Roca (0000-0003-3936-535X) Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 5.1.2.1.2 Explain what access or restrictions will apply to the data? Open The dataset is openly accessible through Zenodo under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, which permits unrestricted reuse, redistribution, and adaptation provided proper credit is given. There are no access restrictions, and the dataset contains no sensitive information requiring protection. 5.1.2.2 Intellectual property rights 5.1.2.2.1 Explain which intellectual property and how will they be dealt with Copyright The dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. This license permits reuse, redistribution, and adaptation of the dataset, provided proper attribution is given. By using CC BY 4.0, copyright permissions and obligations are clearly defined, ensuring open access while retaining recognition of authorship. 5.1.3 Ethical issues 5.1.3.1 What ethical issues and codes of conduct are there, and how will they be taken into account? Other No ethical issues are associated with this dataset, as it contains only synthetic simulation data. The dataset ensures transparency, integrity, and compliance with intellectual property rights. All data are generated in controlled simulations, with no personal or sensitive information, guaranteeing privacy and ethical compliance. The dataset is publicly available to all, promoting equal access regardless of race, gender, or socioeconomic status. 6 Data Sharing And Long-term Preservation 6.1 Data Sharing And Long-term Preservation 6.1.1 How and when will data be shared? Are there possible restrictions to data sharing or embargo reasons? 6.1.1.1 Explain how the data will be discoverable and shared Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 Deposit in a FAIR-enabling data repository 6.1.1.2 Outline the plan for data preservation and give information on how long the data will be retained Zenodo provides long-term archival, guaranteeing persistent access and DOI resolution. All files are retained indefinitely. 6.1.1.3 Explain when the data will be made available The dataset is already publicly available and open access via Zenodo at https://doi.org/10.5281/zenodo.17347998. There are no embargoes or access restrictions. 6.1.1.4 Indicate the expected timely release 2025-11-06 6.1.1.5 Will exclusive use of the data be claimed? No 6.1.1.8 Is it necessary to restrict access to certain communities or to apply a data sharing agreement? No 6.1.2 How will data for preservation be selected, and where data will be preserved long-term? 6.1.2.3 Describe the data to be preserved long-term Simulation (e.g., climate modeling data) 6.1.2.4 Indicate where the data will be deposited Zenodo 6.1.2.5 Indicate how the data will be shared Repository 6.1.2.6 Indicate whether potential users need specific tools to access and (re-)use the data. Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 No specialized tools are required to access the data. CSV and JSON files can be accessed using standard software. PKL files require Python with pandas for full DataFrame loading (pandas.read_pickle). 6.1.3 How will the application of a unique and persistent identifier to each data set be ensured? 6.1.3.1 What type of persistent identifier (PID) will be used? DOI 7 Data Management Responsibilities And Resources 7.1 Data Management Responsibilities And Resources 7.1.1 Who will be responsible for data management? 7.1.1.1 Outline the roles and responsibilities for data management/stewardship activities Miguel Casasnovas (0009-0000-4787-730X) Data capture, metadata production, data quality, data archiving, and data sharing. 7.1.2 What resources will be dedicated to data management and ensuring that data will be FAIR (Findable, Accessible, Interoperable, Re-usable)? 7.1.2.1 Explain how the necessary resources to prepare the data for sharing/preservation have been costed in No cost Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025 5. Software Management Powered by Data Management Plan | MLDR: A Machine Learning-driven Radio Interface LICENSE:CC-BY-4.0 DOI: - 04/12/2025