MLSP 2025 DC

Join our challenge discord server for announcements and Q&A!

In wireless communications, the path loss (PL), also known as the large-scale fading coefficient, quantifies signal strength attenuation between a transmitter (Tx) and a receiver (Rx) due to large-scale effects, such as free-space propagation loss, and interactions such as penetration, reflection, and diffraction, of radio waves with the objects and structures in the propagation environment. Many existing or emerging applications in wireless communications explicitly rely on knowledge of the spatial distribution of the PL, so it is paramount to have reliable estimates of it. Some indicative use cases include: User-cell site association, network deployment, fingerprint-based localization, physical-layer security, optimal power control, path planning, and activity detection.

To evaluate PL or other wireless channel characteristics one can use radio propagation models. Existing channel modeling techniques exhibit a notorious trade-off: deterministic models, e.g., ray tracing, are highly accurate, yet computationally demanding, while the opposite holds for empirical and stochastic models. Recently, substantial effort has been made to develop data-driven methods, which can be trained to yield commensurate accuracy with deterministic propagation models, accompanied by impressive computational efficiency due to the native graph processing unit (GPU) parallelization of deep neural networks. However, while previous research has focused mostly on deep learning (DL)-based PL inference for outdoor urban environments, due to the wide deployment of indoor wireless networks in fifth-generation and beyond (5G/B5G) networks, it is necessary to develop models tailored for indoor environments. In such cases, the refracted electromagnetic field components through obstacles play a more significant role in radio signal propagation, as opposed to the outdoor scenarios that are dominated by reflected field components. Therefore, accurate indoor radio map estimation requires accounting for the greater variety of construction materials and their electromagnetic properties.

To advocate further research in this direction and facilitate fair comparisons in the development of DL-based radio propagation models in the less explored case of PL radio maps in indoor environments, and motivated by the success of the First Pathloss Radio Map Prediction Challenge at ICASSP 2023 and the First Indoor Pathloss Radio Map Prediction Challenge at ICASSP 2025, we share an indoor PL radio map dataset generated through ray tracing simulations, launching the Sampling-Assisted Pathloss Radio Map Prediction Data Competition at MLSP 2025. The challenge consists of two tasks that explore the prediction of PL radio maps using DL-based methods when aided by ground truth PL samples with varying sampling rates from the propagation environment. The first task considers a uniform random selection of sampling locations, while the second task focuses on investigating the impact of PL sampling location selection, thus requiring the exploration of efficient and effective sampling techniques to improve the fidelity of radio map estimation over random sampling.

The 5 teams with the best overall performance in these two tasks will be invited to submit a paper and present it at the IEEE International Workshop on Machine Learning for Signal Processing (MLSP) 2025, and the accepted papers will be published in the MLSP proceedings.

The deadline for submitting trained models, test codes and the radio map estimates on the test dataset is May 10, 2025.

Support on the dataset and the instructions will be provided by the organizing team.

IMPORTANT NOTE: The intellectual property (IP) of the shared/submitted material (e.g. code) will not be transferred to the challenge organizers. When such material is made publicly available by a participant, an appropriate license should accompany.