pdfs describe variation of channel loss. Some of them are exact, but usually only for simple or well specified geometries or situations, for which they can be derived. In other cases they may be empirical, in that they are close matches to what has been observed. Presumably, machine learning can be used to measure the actual variation of channel loss and generate a specific model for each channel that might usually be better than a theoretical ideal or empirical approximation, and take account of parameters like time of day, for instance.

Dear Ajay Yadav

These articles might be an asset, have a look:

1. https://liu.diva-portal.org/smash/get/diva2:1442845/FULLTEXT01.pdf

2. https://arxiv.org/pdf/1503.00877.pdf

3. https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8660712

4. https://www.researchgate.net/publication/220144678_Worse-than-Rayleigh_Fading_Experimental_Results_and_Theoretical_Models

5. https://lib.dr.iastate.edu/cgi/viewcontent.cgi?article=2267&context=rtd&unstamped=1

6. https://www.mdpi.com/2073-8994/12/3/389/htm

7. https://www.hindawi.com/journals/wcmc/2018/6497340/

8. https://www.hindawi.com/journals/wcmc/2021/8838792/

Kind Regards

Qamar Ul Islam

The common pathloss models are essentially obtained by machine learning. People make extensive pathloss measurements in certain environment (e.g., rural or urban) and then they try to fit a statistical distribution to the results, so one can generate almost identically distributed pathloss realizations without having to make new measurements.

Rayleigh, lognormal, and Rician are just names of the statistical distribution that considered. The nice things with these distributions is that you need less than 5 parameters to describe the results pathloss models.

One can certainly create more intricate models with many more parameters, but then the risk of overfitting to measurement data is high. Statistical models will always be less accurate than real measurements or ray-tracing, but it serves a different purpose. It is like quantitative channel model vs. quantitative channel model.

In summary, when you say "Channel modeling using machine learning", you need to think of what it is that you want to achieve. What do you want your model to learn? How can you ensure generalizability of the results?

What caught my attention the most from the comment of Prof Emil is how he is relating statistical modelling and machine learning. The part I gave below quite interesting and needs further thoughts.

Do I understand correct Prof Emil:

We have measurements and assume a state-of-the-art distribution to model it and then predict new data based on the assumed underlying distribution, then this does not become machine learning, since we require the underlying distribution in advance.

However, if we build our own distribution model based on the measurements and predict accordingly, then this becomes similar to what machine learning does, because we deveop a model from the data (e.g., learning) as in machine learning.

Did I get it what you said correctly?

"The common pathloss models are essentially obtained by machine learning. People make extensive pathloss measurements in certain environment (e.g., rural or urban) and then they try to fit a statistical distribution to the results, so one can generate almost identically distributed pathloss realizations without having to make new measurements."

I think that machine learning will be better, but a huge of measurements must be carried out and this measurement data will be used for training a reliable channel model such that this model can be used for channel coefficients generation.

Thank you all.