Originally posted by MossyRocks: Actually by averaging noise images you are actually driving down the truly random noise portion and this obeys the 1/sqrt(# shots) rule. The only way to get rid of that is as you say to collect more light frames. Getting rid of the random noise in other shots like darks and bias frames is also desirable so that you can clearly see the systematic noise, or probably more correctly call errors, from various components. The lower the errors of these components the more shots you need to move the true noise floor below these.
To be very clear (because I'm not sure it will be to anyone else reading this) averaging dark frames only reduces the amount of random shot noise that is introduced by using dark frames in the first place. It does NOT reduce the random shot noise caused by the dark signal that is in the light frames. Only more total exposure can do that. (Note that the non-random component of the dark signal is actually mainly suppressed in all but the very oldest cameras, to varying degrees of success.) Use of a master dark frame WILL add some random shot noise. But if you averaged a lot of dark frames to make your master dark, it will be small. Similarly, using a master bias frame will not reduce the random component of read noise that is in each of the light frames. In fact, it will introduce more (but again very small if you averaged a lot of bias frames). A master bias will only reduce the non-random component of the read-noise signal.
Hence, using dark frames and bias frames DOES add extra noise to the image, the amount depending on how many frames were averaged for each. But this hopefully small amount of extra noise is better than leaving FPN, which may include amp glow, in the image, which is what dark frames and bias frames (if there is a non-random component to the read noise) are designed to eliminate. Dark frames also get rid of hot pixels. But there are several alternative methods of getting rid of hot pixels including dithering. (Dithering can also reduce the magnitude of FPN to the point of insignificance if it is not too severe.) Those alternative methods do not add any extra random noise.
All of that said, if the sub-exposures are exposed properly the read noise will be insignificant compared to the overall signal (sky background + objects). And in the presence of even a moderate amount of light pollution, the random noise introduced by the sky background will swamp even the random thermal noise from most of today's cameras on all but the warmest nights. So, random read noise and for many people also random thermal noise are the least of the random noises one needs to worry about.
(Some may object to my putting "random" in front of noise. But I do that because some people also define non-random components of a signal as noise if those non-random components are unwanted. I find that usage makes discussion harder. However, given that usage I want to always make it clear when I am talking about random (shot) noise.)