Algorithm 1 Proposed DL Channel Estimation scheme

1.Initialize the learnable parameters of an LSTM layer (W, R, b), where W is the input weights, R is the recurrent weights, and b is the bias.   Inputs ▪Number of Iterations. ▪The distance of each user from the BS. ▪Path loss exponent. ▪Generate Initial random Rayleigh channel coefficients for each user based on the channel model. ▪Generate known pilot symbols. ▪Assign the initial power factors for each user. ▪Identify the size of training sequence $({\mathit{L}}_{\mathit{T}})$ and size of testing sequence $\left({\mathit{L}}_{\mathit{S}}\right)$.   Procedure 2.Assign the training sequence $({\mathit{Z}}_{\mathit{T}})$, testing sequence $({\mathit{Z}}_{\mathit{S}})$, and the desired coefficients $({\mathit{Z}}_{\mathit{D}})$. 3.Calculate the mean and variance $({\mathit{\mu }}_{\mathit{T}}, {\mathit{\sigma }}_{\mathit{T}}^{\mathbf{2}})$ for channel coefficients in training sequence at each iteration. 4.Normalizing the training data ${\mathit{Z}}_{\mathit{T}}\to {\mathit{Z}}_{\mathit{N}\mathit{T}}$ based on $({\mathit{\mu }}_{\mathit{T}}, {\mathit{\sigma }}_{\mathit{T}}^{\mathbf{2}})$. 5.Characterize the relationship between consecutive normalized training sequences as ${\mathit{Z}}_{\mathit{N}\mathit{T}\approx }({\mathit{X}}_{\mathit{N}\mathit{T}}, {\mathit{Y}}_{\mathit{N}\mathit{T}})$. 6.Initialize the training network $({\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}})$ and assign the following: ▪Number of layers. ▪Number of hidden units. ▪Training parameters. 7.Use $({\mathit{X}}_{\mathit{N}\mathit{T}}, {\mathit{Y}}_{\mathit{N}\mathit{T}})$ as inputs for the training model $({\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}}$). 8.Predict the output normalized coefficients $({\mathit{Y}}_{\mathit{N}\mathit{P}})$. 9.Update the state of training model $({\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}}$).    For i = 1: ${\mathit{L}}_{\mathit{T}}$    $[{\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}}, {\mathit{Y}}_{\mathit{N}\mathit{P}}]$ = Predict and Update state $({\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}}, {\mathit{Z}}_{\mathit{N}\mathit{T}})$    End 10.Denormalize ${\mathit{Y}}_{\mathit{N}\mathit{P}}\to {\mathit{Y}}_{\mathit{P}}$. 11.Calculate RMSE $({\mathit{Z}}_{\mathit{D}}-{\mathit{Y}}_{\mathit{P}})$ and Loss function. 12.Update the state of $({\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}} )$ and reset the values for ${\mathit{Y}}_{\mathit{P}}$. 13.Normalize testing data ${\mathit{Z}}_{\mathit{S}}\to {\mathit{Z}}_{\mathit{N}\mathit{S}}$, using $({\mathit{\mu }}_{\mathit{T}}, {\mathit{\sigma }}_{\mathit{T}}^{\mathbf{2}})$. 14.Use normalized testing data $\left({\mathit{Z}}_{\mathit{N}\mathit{S}}\right)$ as inputs for updated trained network $({\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}}$).     For i = 1: ${\mathit{L}}_{\mathit{S}}$     $[{\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}}, {\mathit{Y}}_{\mathit{N}\mathit{P}}]$ = Predict and Update state $({\mathit{T}}_{\mathit{n}\mathit{e}\mathit{t}}, {\mathit{Z}}_{\mathit{N}\mathit{S}}$)     End     Outputs 15.Predicted normalized channel coefficients ${\mathit{Y}}_{\mathit{N}\mathit{P}}$. 16.Denormalize ${\mathit{Y}}_{\mathit{N}\mathit{P}}\to {\mathit{Y}}_{\mathit{P}}$. 17.Calculate RMSE $({\mathit{Z}}_{\mathit{D}}-{\mathit{Y}}_{\mathit{P}})$ & Loss function. 18.Estimate channel taps (trained DNN model, pilot symbols).