This challenge is promoted by Universitat Pompeu Fabra (UPF) and is part of the ITU Artificial Intelligence/Machine Learning in 5G Challenge. Further participation details can be found in ITU AI/ML 5G Challenge: Participation Guidelines. Participation is open to ITU members and any individual from an ITU Member State. “Participants” are individuals or companies that participate in the ITU AI/ML in 5G Challenge, providing solutions to problem sets of the Challenge. This challenge is open to the two categories of participants: student and professional.

Additional resources:

  • Video introducing the problem statement and the dataset
  • Presentation (in pdf) introducing the problem statement and the dataset


  1. Registration deadline [new!]: 30 June 2020 21 August 2020 (see ITU-T guidelines)

  2. Global Round duration: June - October 2020

  3. Training & Validation data sets: Available now

  4. Test data set: [to be defined]

  5. Deadline to submit a solution: 15 September 2020

  6. Announcement of the winners: October 2020

How to participate?

  1. If you don't have an ITU account, please follow this guide to create one for challenge registration.
  2. Register on ITU AI/ML in 5G challenge website with your ITU account.
  3. Fill out the ITU AI/ML in 5G Challenge Participants Survey​​ to select problem statement ITU-ML5G-PS-013. You can enroll as a team with 1-4 members.  
  4. Begin to work on this problem and submit your results. We will begin to accept submissions from July 31, 2020 and the submission deadline is September 15, 2020.


Next-generation IEEE 802.11 wireless local area networks (WLANs) are called to face the challenge of providing high performance under complex situations, e.g., to provide high throughput in massively crowded deployments where multiple devices coexist within the same area. To address the performance challenges posed by the requirements derived from novel use cases, one of the features receiving more attention is Channel Bonding (CB) [1, 2], whereby multiple frequency channels can be bonded with the aim of increasing the bandwidth of a given transmission, thus potentially improving the throughput.

Since its introduction to the 802.11n amendment, where up to two basic channels of 20 MHz could be bonded to form a single one of 40 MHz, the specification on CB has evolved and currently allows for channel widths of 160 MHz. However, using wider channels entails spreading the transmit power over the selected channel width, which can potentially affect the data rate used for the transmission, and the capabilities of the receiver on decoding data successfully. Moreover, the potential gains of CB in crowded deployments is further hindered because of the multiple inter-BSS interactions, which may provoke contention based on the global channel scheme (i.e., the set of channels allocated to each BSS). In particular, CB in dense deployments is a problem with a combinatorial action space.

Problem statement and potential applications in science

To address the CB problem in WLANs, we propose the application of Machine Learning (e.g., Deep Learning) to predict the performance of an OBSS where different combinations of channel schemes are allocated to the different BSSs. The main purpose is, therefore, to predict the throughput that a BSS would obtain according to the data extracted from simulated deployments generated based on different random parameters, including channel allocation, location of nodes, and number of STAs per BSS.

Performance prediction in WLANs can be used to optimize the planning phase of a given deployment or improve the performance during the operation of a WLAN.

Data set

The assets provided comprise both training and validation data sets of CB in IEEE 802.11 WLANs. Each of the data set partitions includes two different enterprise-like scenarios, where a different fixed number of BSSs coexist in the same area. Scenario 1 is composed of 12 APs, each one with 10 to 20 associated STAs, while Scenario 2 has 8 APs with 5 to 10 STAs per AP. Each of the scenarios is reproduced in random deployments for three different map sizes. The summary of the provided deployments is as follows:

  • Scenario 1 (12 APs, 10-20 STAs):

    • Scenario 1a (map width = 80 x 60 m): 50 random deployments

    • Scenario 1b (map width = 70 x 50 m): 50 random deployments

    • Scenario 1c (map width = 60 x 40 m): 50 random deployments

  • Scenario 2 (8 APs, 5-10 STAs):

    • Scenario 2a (map width = 60 x 40 m): 50 random deployments

    • Scenario 2b (map width = 50 x 30 m): 50 random deployments

    • Scenario 2c (map width = 40 x 20 m): 50 random deployments

The data set can be downloaded here.

Overview of the data set

Two types of files are provided: 

  1. Nodes' input files: input files defining the network deployments used to run the simulator
  2. Output Komondor: output generated by the simulator after running the nodes' input file

To train an ML model, participants should use both input features (deployment characteristics) and labels (performance). The features can be obtained from both nodes' input files and Komondor's output. Apart from the performance labels (throughput and airtime), the output of the simulator provides the RSSI list and the interference map (further described below).

Input features:

  • Deployment characteristics (provided in input nodes files): labels and locations of the nodes and the selected channel scheme. Input files contain other static information (e.g., CW size) that is not recommended to be used for training.

  • RSSI list (provided in output files): RSSI in dBm that each device receives from its AP. Each row represents a BSS.

  • Interference map (provided in output files): inter-BSS interference sensed from each AP in dBm. Each row represents the signal strength that each AP receives from all the other APs.

Output labels:

  • Per-STA throughput (provided in output files): average throughput experienced by each STA at the end of the simulation, being the throughput of the AP the aggregate throughput of the BSS (i.e., the sum of all the individual throughput allocated to each STA in the BSS).

  • Airtime per channel (provided in output files): percentage of time each BSS occupies each of the assigned channels. E.g., if a given BSS transmits during the 80% of the time in both channels 3 and 4, the airtime will be provided as {80, 80}. Important: the airtime is an auxiliary label that seeks to provide a further understanding of the interactions taking place in each simulated deployment.

Example of data set entry

For the sake of illustration, here we show a simplified data set entry.

  • Input node files (contains the characteristics of the deployment). The most important features are shown in the following table.


node_type (0:AP, 1:STA)

wlan_ code











































* Notice that input files include other static information such as the channel bonding model, the minimum Contention Window (CW), etc. This information is recommended to be ignored for training your algorithm.

  • Simulation output files (includes labels and features). Information is provided in the following order:

- Per-STA throughput (in Mbps): {100, 100, 75, 75}

* In this case, there is only one STA per BSS, so the aggregate throughput is the same as the throughput obtained by STAs. The order of appearance matches the indicated in the nodes' input files. The size of this array is [#APs + #STAs].

- Airtime per BSS per channel (in %): {80, 80, 80, 50, 50, 50, 50, 50; 50, 50, 50}

* Note that information is shown for each BSS (separated by a semicolon) and only for the set of selected channels. Thus, the size of this array is dynamic.

- RSSI list (in dBm): {Inf, -67, Inf, -72}

* Notice that APs are also included in the list, but the RSSI received from themselves is marked as ‘Inf’. The size of this array is [#APs, #STAs].

- Interference map (in dBm):

{Inf, -87;

-87, Inf}

* Notice that each AP receives an Infinite amount of power from itself. The size of this array is [#APs, #APs].

The output results for all the deployments of each scenario is provided in the same file (e.g., “script_output_sce1a.txt”). Each deployment is introduced with the following header:

KOMONDOR SIMULATION ‘sim_input_nodes_sce1a_deployment00.csv’ (seed 1992),

which includes the name of the input nodes file used to conduct the simulation (in this case, ‘sim_input_nodes_sce1a_deployment00.csv’) and the random seed.

How was the data set generated?

To generate the data set, we have used the Komondor simulator (, Commit: 93063aafe6d62991da680262fb6735b5d9b056f4), an IEEE 802.11ax-oriented network simulator developed at the UPF. Komondor has been validated against ns-3 [3] and includes novel functionalities such as channel bonding and spatial reuse.

The main parameters used to conduct the simulations are as follows:

  • Simulation duration: 10 seconds
  • Downlink UDP traffic (full buffer)
  • The traffic of a BSS is uniformly spread among STAs
  • Random channel allocation
  • Random number of STAs per BSS
  • Location of APs is fixed at the center of each cell
  • The location of STAs is randomly selected around the AP
  • A dynamic channel bonding policy has been applied, whereby nodes attempt to transmit over the widest possible channel

Evaluation criteria

  • Participants must use the provided data set to train a machine learning algorithm.

  • The output of the ML algorithm should be able to predict the performance obtained in a new network deployment. In particular, the expected throughput of each BSS in the deployment must be provided.

  • The choice of the ML approach is decided by each participant (neural network, linear regression, decision tree, etc.).

  • A test data set will be provided to evaluate the performance of the proposed algorithms.

  • The evaluation of the proposed algorithms will be based on the average squared-root error obtained along with all the predictions compared to the actual result in each type of deployment.

  • The winners will be invited to publish the results in an academic publication.


Francesc Wilhelmi ([email protected])

Boris Bellalta ([email protected])

ITU AI Challenge committee ([email protected]​​)


[1] Barrachina-Muñoz, S., Wilhelmi, F., & Bellalta, B. (2019). Dynamic channel bonding in spatially distributed high-density WLANs. IEEE Transactions on Mobile Computing.

[2] Barrachina-Muñoz, S., Wilhelmi, F., & Bellalta, B. (2019). To overlap or not to overlap: Enabling channel bonding in high-density WLANs. Computer Networks, 152, 40-53.

[3] Barrachina-Muñoz, S., Wilhelmi, F., Selinis, I., & Bellalta, B. (2019, April). Komondor: a wireless network simulator for next-generation high-density WLANs. In 2019 Wireless Days (WD) (pp. 1-8). IEEE.