This is a classical deconvolution problem in that your received signal consists of multiple delayed and scaled copies of your transmitted signal, no different than what could be modeled with an FIR filter (which is the sum of delayed and scaled copies of the input signal), and your goal is to reconstruct the original transmitted signal from the received signal.
The received signal is the convolution of transmitted signal with the effective channel given by the "filter" described above which in this case would be the multiple different reflections from the surface of the wall. So we have $r[n] = t[n]\ast h[n]$ (received signal $r[n]$ is the convolution of the transmitted signal $t[n]$ and the channel $h[n]$), with two knowns $r[n]$ and $t[n]$ and we wish to solve for $h[n]$. This is a classical deconvolution problem typically solved using the Wiener-Hopf equations resulting in a least squared solution when there are sufficient samples for an overdetermined equation. This is explained in detail along with MATLAB code at this post here which was specific to an audio signal but would equally apply to the OP's case of a broadband microwave signal from 3.2 GHz to 4.8 GHz as he explained in comments:
Compensating Loudspeaker frequency response in an audio signal
As described above this would provide for a solution to the actual channel modeled as an FIR filter, but the process can equally be used to solve for an equalizer $c[n]$ which would optimally in a least squared sense return the received signal to the closest copy of what was transmitted simply by swapping $t[n]$ and $r[n]$ since in the receiver, the copy of the transmitted signal $t[n]$ is $r[n]\ast c[n]$.
Some precautions when using "sounding" signals to determine a channel:
The transmitted signal would ideally be white over the frequency band of interest. Certainly if there is any frequency portion with no signal content, no solution can be determined for that portion of frequency (since there would be no input/output relationship). If the signal is not white, the least squared solution would be weighted according to the respective spectral density over the entire bandwidth of interest.
If the delay spread is long compared to the inverse of the bandwidth (so becomes a greater problem for wideband signals), deep nulls can occur in the received spectrum due to the cancellation effects of the various delay paths. This is known as "Frequency Selective Fading" and makes equalization and deconvolution challenging as noise enhancement would occur at the locations where deep spectral nulls exist. The received spectrum should be reviewed to confirm that such deep nulls do not exist. For communication systems this is typical motivation for use of non-linear equalizers such as decision-feedback equalizers which would perform better for equalization under such conditions. (see non linear equalizer vs linear equalizer)
In this related post I showed how the channel estimation formulas can be used to determine the group delay of a channel:
How determine the delay in my signal practically