SIA: Symbolic Interpretability for Anticipatory Deep Reinforcement Learning in Network Control

This repository contains the code and resources for the research paper accepted at IEEE INFOCOM 2026:

Paper preprint: https://arxiv.org/abs/2601.22044

SIA: Symbolic Interpretability for Anticipatory Deep Reinforcement Learning in Network Control
MohammadErfan Jabbari, Abhishek Duttagupta, Claudio Fiandrino, Leonardo Bonati, Salvatore D'Oro, Michele Polese, Marco Fiore, and Tommaso Melodia

IMDEA Networks Institute, Spain
Universidad Carlos III de Madrid, Spain
Northeastern University, Boston, USA

Email: {name.surname}@networks.imdea.org, {l.bonati, s.doro, m.polese, t.melodia}@northeastern.edu

Abstract

Deep Reinforcement Learning (DRL) promises adaptive control for future mobile networks, but conventional agents remain reactive: they act on past and current measurements and cannot leverage short-term forecasts of exogenous KPIs such as bandwidth. Augmenting agents with predictions can overcome this temporal myopia, yet uptake in networking is scarce because forecast-aware agents behave as closed boxes; operators cannot tell whether predictions truly guide decisions or justify the added complexity. We propose SIA, the first interpreter that exposes in real time how forecast-augmented DRL agents operate. SIA combines symbolic AI abstractions with per-KPI knowledge graphs to generate explanations and introduces a new influence score metric. SIA achieves sub-millisecond runtime, more than 200× faster than existing XAI methods. We evaluate SIA on three diverse networking use cases, uncovering hidden issues including temporal misalignment in forecast integration and reward-design biases that trigger counterproductive policies. These insights enable targeted fixes: a redesigned agent improves average video bitrate by 9%, and SIA's online Action-Refinement module improves RAN-slicing reward by 25% without retraining. By making anticipatory DRL transparent and tunable, SIA lowers the barrier to proactive control in next-generation mobile networks.

Citation

If you find this work useful, please cite our paper:

@inproceedings{sia2026,
  title={SIA: Symbolic Interpretability for Anticipatory Deep Reinforcement Learning in Network Control},
  author={Jabbari, MohammadErfan and Duttagupta, Abhishek and Fiandrino, Claudio and Bonati, Leonardo and D'Oro, Salvatore and Polese, Michele and Fiore, Marco and Melodia, Tommaso},
  booktitle={IEEE INFOCOM 2026},
  year={2026},
  note={DOI to be added after publication. Available online: https://github.com/RAINet-Lab/SIA}
}

This repository contains the public SIA implementation and artifacts for three representative networking workloads: adaptive bitrate streaming (ABR), massive MIMO scheduling, and RAN slicing. It includes the migrated SIA core package, forecasting utilities, curated notebooks, and use-case-specific analysis code used to reproduce the public parts of the paper workflow.

Key Contributions

• Introduces SIA, a symbolic interpretability framework for anticipatory DRL agents in mobile networks.
• Provides a reusable core package under src/sia/ for symbolic abstraction, decision-graph construction, and forecasting support.
• Validates SIA across three networking domains: ABR, massive MIMO, and RAN slicing.
• Includes a public ABR bootstrap pipeline that reconstructs the trace layout expected by the migrated agents from publicly accessible network traces.
• Documents the data-access boundary for the confidential or third-party datasets used in the MIMO and RAN-slicing workflows.

Repository Structure

SIA/
├── src/sia/                  # Reusable SIA package
│   ├── core/                 # Symbolization, knowledge-graph, and decision logic
│   └── forecasting/          # Forecasting layers and support code
├── use_cases/
│   ├── abr/                  # ABR agents: vanilla, Lumos, Xatu, and SIA refiner
│   ├── mimo/                 # Massive-MIMO notebooks and analysis entrypoints
│   └── ran_slicing/          # RAN-slicing notebooks and public result workflow
├── notebooks/                # Curated top-level entrypoint notebooks
├── scripts/                  # Bootstrap helpers and smoke tests
├── docs/                     # Architecture and migration notes
├── data/                     # Lightweight placeholders for raw/processed data roots
└── tests/                    # Lightweight repository checks

Getting Started

1. Clone the repository

   git clone git@github.com:RAINet-Lab/SIA.git
   cd SIA

2. Create an environment and install dependencies

   python -m venv .venv
   source .venv/bin/activate
   pip install -e .

3. Run the lightweight validation checks

   python scripts/smoke_imports.py
   python tests/test_repo_layout.py

4. Bootstrap the public ABR traces

   python scripts/bootstrap_public_data.py --dataset abr

Data Access

• ABR: Publicly bootstrapable through scripts/bootstrap_public_data.py, which downloads Norway mobility traces and reconstructs the layout expected by the migrated ABR code.
• RAN slicing: The dataset is confidential. This repository ships the scripts, notebooks, and public results only. For dataset access, contact the Northeastern University coauthors via the paper email block: {l.bonati, s.doro, m.polese, t.melodia}@northeastern.edu.
• Massive MIMO: This repository ships the analysis code and notebooks, but not the original training traces. The environment is based on A Deep Reinforcement Learning-Based Resource Scheduler for Massive MIMO Networks (IEEE TMLCN 2023). For trace access, contact the original paper authors, especially the Rice University group behind that scheduler environment.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Contact

For questions, please open an issue on this repository or contact the paper authors.