The growing complexity of cellular networks demands smarter optimization methods to balance coverage, capacity, and user demand. Data-driven methods, which operate on direct feedback from monitoring data, show promise but struggle with scalability, safety, and sample efficiency. In contrast, model-based schemes enable efficient optimization but suffer from potential model misspecification.
The Christian Doppler Laboratory for Digital-Twin Assisted AI for Sustainable Radio Networks addresses these challenges through a hybrid approach that combines the strengths of both model-free and model-based domains. This is achieved by integrating a Monte Carlo Tree Search (MCTS)-based agent with adifferentiable network twin. The twin guides exploration through fast and scalable gradient-based optimization, ensuring both safety and efficiency in network configuration.
To foster collaboration, we are releasing the differentiable throughput model, a critical component of the network twin. This model integrates real-world monitoring data and supports gradient-based optimization, enabling load-aware tuning of key network parameters such as transmit power and antenna tilt. The model is available https://squid.nt.tuwien.ac.at/gitlab/leller/differentiable_throughput_model.
For more details, please refer to:
As the demand for wireless connectivity grows, especially with the upcoming 6G networks, mobile network operators must plan deployments carefully to manage capacity, reduce interference, and optimize energy use—particularly in complex urban environments. While traditional ray tracing methods are accurate, they are computationally expensive. Data-driven machine learning models offer a faster, efficient alternative using real-world data for signal prediction.
The Christian Doppler Laboratory for Digital-Twin Assisted AI for Sustainable Radio Networks addresses these challenges by developing machine learning models trained on drive-test and crowdsourced MDT data. These models not only predict signal strength but also provide uncertainty estimates to enable accurate and reliable network planning.
We are now making these trained models publicly available, along with a minimal working example. Access the models https://squid.nt.tuwien.ac.at/gitlab/leller/ieee_access_deep_learning_network_planner to benchmark and compare for your own research.
For more information, refer to our studies on: