How to fix the random seed of sample function in R - r

Question

I am wondering how to fix the random seed of sample function in R.
An easy example is here:
set.seed(1)
tmp <- c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
sample(tmp, size = 3) # (*)
#[1] 9 4 7
sample(tmp, size = 3) # (**)
#[1] 1 2 5
sample(tmp, size = 3) # (***)
#[1] 7 2 3
sample(tmp, size = 3) # (****)
#[1] 3 1 5
# retry set the random seed
set.seed(1)
sample(tmp, size = 3) # (*)
#[1] 9 4 7
sample(tmp, size = 3) # (**)
#[1] 1 2 5
sample(tmp, size = 3) # (***)
#[1] 7 2 3
sample(tmp, size = 3) # (****)
#[1] 3 1 5
Then, when I repeat the sample function, the returned values change, although pattern of the change is uniform.
I cannot understand why sample function is differently affected from the set.seed.
How can I fix the return from sample function?
Thank you for your time to spend to read this question.

Answer 1

something like this may be
{set.seed(1); sample( tmp, 3 )}
[1] 9 4 7
This will return same result whenever called

Answer 2

I think your post sort of answers the question. If you notice the first sample after setting the seed are the same, which is what it's expected to do:
set.seed(1)
tmp <- c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
sample(tmp, size = 3)
[1] 9 4 7
sample(tmp, size = 3)
[1] 1 2 5
set.seed(1)
tmp <- c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
sample(tmp, size = 3)
[1] 9 4 7
sample(tmp, size = 3)
[1] 1 2 5
The first set of three draws after setting the seed are the same. The second draw of three is the same.
Perhaps it could be easier to see how set.seed works by looking at the first draw of 6 random selected values:
set.seed(1)
tmp <- c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
sample(tmp, size = 6)
[1] 9 4 7 1 2 5

Answer 3

Your question seems to reflect a basic misunderstanding of Pseudo-Random Number Generators (PRNGs) and seeds. PRNGs maintain an internal state containing a fixed number of bits, and advance from the current state to the next state by applying a deterministic function. That function determines how "random" the sequence of state-based outputs appears to be, even though it's not actually random (hence the pseudo prefix). Given a fixed number of bits, the algorithm will eventually repeat its internal state, at which point the entire sequence repeats since it's deterministic. How long it takes to do this is called the cycle length of the PRNG.
Once you realize that all PRNGs will eventually repeat their sequence, it's easy to see that a seed does nothing more (nor less) than provide an entry point to that cycle. A lot of people mistakenly think the seed influences the quality of the PRNG, but that's incorrect. The quality is based on the deterministic transition function, not on the seed.
So why do PRNGs allow you to specify a seed? Basically it's for reproducibility. Firstly, it's really hard to debug a program if you can't reproduce the data which exposed your bug. Re-running with the same seed will allow you to trace through to whatever caused a problem. Secondly, this allows for fairer comparisons of alternative systems. Let's say you want to compare how a grocery store, or a bank, works when you add an additional clerk or teller. If you run both configurations with the same set of customers and customer transactions by controlling seeding, the differences you see between those configurations are directly attributable to the change in the system configuration. Without controlling the seeding, differences may be damped or amplified by the randomness of the sequence. Yes, in the long run that would come out by taking a larger sample, but the point is that you can reduce the variance of the answer by clever use of seeding rather than boosting the sample size.
Bottom line - your example is working exactly as expected. Resetting the seed produces the same sequence of randomness when sampled in the same way.