Follow-up on Dataset Generation and Channel Scaling

[Marwanmohamedkher]

Dear Dr. Zhu,

Thank you very much for your previous response and for outlining the dataset generation procedure. It was very helpful in guiding my own implementation.

Following your description, I implemented the dataset generation pipeline by:

Randomly generating 4 user locations within a 5-meter radius around (100,15),

Applying the LoS/NLoS path loss models from Table II of your GMML paper,

Generating Rician fading channels (with κ = 10) for both the BS–RIS and RIS–User links,

Using the standard carrier frequency (28 GHz), 64 BS antennas, and 100 RIS elements.

After this, I observed that the generated channel matrices had significantly lower average power compared to the provided dataset.mat file. For instance:

The average power of G (BS–RIS) was around –71 dBm,

The average power of H (RIS–Users) was around –73 dBm.

To reconcile this difference and enable fair benchmarking (e.g., for training GMML ), I applied a constant scaling factor to match the average power of the original dataset:

scaling_factor_G = sqrt(mean(abs(G_ref(:)).^2) / mean(abs(G_generated(:)).^2));
scaling_factor_H = sqrt(mean(abs(H_ref(:)).^2) / mean(abs(H_generated(:)).^2));

I’d like to confirm:

Did you apply a similar power normalization or scaling in your original data?

Is applying such a scaling factor a reasonable approach to align with your dataset for fair evaluation?

I appreciate your time and guidance, and thank you again for the valuable contribution your work provides to the community.

Best regards,
Marwan Mohamed