[JOSS] make functions run deterministic

[matbesancon]

Following the convention from Julia stdlib's Random and various packages from the JuliaStats ecosystem, every function which performs a random sampling should take in parameter the Random Number Generator (which is usually a subtype of Random.AbstractRNG)

[kirui93]

@matbesancon Is this really necessary for functions like rand() and randn()?

[matbesancon]

That's precisely for these functions. You would take a RNG in all functions and pass it to the rand and randn calls

[kirui93]

Yes. I have seen what you mean. This means that I need to use the package Random for this. Thank you. I will implement immediately.

[matbesancon]

Linking to the review: https://github.com/openjournals/joss-reviews/issues/1912

[kirui93]

@matbesancon I will close this issue then because I have already included using Random RNG in the package through MersenneTwister.

[kirui93]

According to Julia's documentation, every function which performs a random sampling should take in a parameter the Random Number Generator (which is usually a subtype of Random.AbstractRNG. Also, random number generation uses the Mersenne Twister library via MersenneTwister objects. We employed this idea and included this functionality for all the parts of the code where we generate random numbers.

[matbesancon]

https://github.com/kirui93/ScenTrees.jl/blob/master/src/StochPaths.jl

This is not quite right, the rng as a global variable makes each run depend on the number of calls of a function. The rng should be an argument passed to functions instead

[kirui93]

https://docs.julialang.org/en/v1/stdlib/Random/

I used the concept in the above julia documentation. Another option was to use Random.seed!().

[kirui93]

Both MersenneTwister and Random.seed!() works in a similar manner. Consider the below,

julia> MersenneTwister(01102019) == Random.seed!(01102019) true

[kirui93]

In https://github.com/kirui93/ScenTrees.jl/blob/master/src/StochPaths.jl, rng = MersenneTwister(0101219) and I have passed it as an argument of randn() in all the functions.

[kirui93]

Hello @matbesancon, do you suggest using as in the following example instead of the above?

julia> rand(rng::AbstractRNG, 3,2)
3×2 Array{Float64,2}:
 0.496169  0.449182 
 0.732     0.875096 
 0.299058  0.0462887

[matbesancon]

In any function that needs rng, it should be an argument of this function:

GaussianSamplePath1D()

# replace with 
GaussianSamplePath1D(rng)

[kirui93]

For the stochastic approximation process, we need a function that doesn't take any inputs. That is why we coded these examples without any inputs. This is why we preferred to have rng as global variable.

[matbesancon]

If at some point you need the function to be run randomly, you can always call it with:

GaussianSamplePath1D(Random.GLOBAL_RNG)

while if it is defined without argument, making it deterministic is more painful, since one needs to set the seed outside of it

[kirui93]

@matbesancon These functions that you have pointed out were just examples of the functions that you can pass to the stochastic approximation process. The main factor that we are insisting is that these kind of functions should not take any inputs and that is why I prefer to have these examples in the current state. Having the seed outside is really expensive but we prefer to have it that way to be in the scope of the rules of the function that we are passing to the approximating functions.

[kirui93]

Hello @matbesancon Any comments so far? Thank you.

[matbesancon]

The main factor that we are insisting is that these kind of functions should not take any inputs

Is there a reason for that? I fail to see why. In my example above, the function is taking the RNG as a normal argument, but you could also pass is as a keyword argument with a default to Random.GLOBAL_RNG, which means the default would be the same as you are currently having

[kirui93]

Hello @matbesancon , In the functions TreeApproximation! and LatticeApproximation, we pass these kind of functions as an argument and in the process, we keep calling them for every iteration we make. In this case, therefore, our requirement was to have a function to be approximated that takes no inputs. The user can use function closures or anything else for these kind of functions as long as he/she will pass a function with no inputs.

[matbesancon]

In the functions TreeApproximation! and LatticeApproximation, we pass these kind of functions as an argument and in the process

It is not a breaking change with your current code, if rng is taken as keyword argument, you can still call the function without any argument and it would work. The difference is that now you can also call it with a controlled input source of randomness

[kirui93]

In the functions TreeApproximation! and LatticeApproximation, we pass these kind of functions as an argument and in the process

It is not a breaking change with your current code, if rng is taken as keyword argument, you can still call the function without any argument and it would work. The difference is that now you can also call it with a controlled input source of randomness

I prefer to have it the way it is now. These functions GaussianSamplePath1D() and the rest are just examples of stochastic processes that we require the user to provide.

[matbesancon]

Ok, I don't think it's the ideal setting, but I will not consider that blocking for the submission to JOSS.

[kirui93]

I will still consider what you have proposed and see how having rng as an input to the functions affects the main functions.

[kirui93]

Let me refer you to this line, for example,

https://github.com/kirui93/ScenTrees.jl/blob/master/src/TreeApprox.jl#L31

if we include the option of using rng for these stochastic functions, then we are forcing the user of the package to input a function which takes rng as an input all the times. This may not be the case for all situations. That is why we prefer a function with no inputs.