Rename ':noise' key in solution storage about sddp.jl HOT 3 CLOSED

Hey Vitor, implementing an out of sample simulator sounds great!

One issue is that renaming the key would break a lot of existing code.

I'd be more in favour a bigger reorganisation that included returning the actual noise object (or implementing the oracles since they seem the next most important thing) so things broke in a big way rather than quietly.

As I think about this more, maybe there should be :noiseidx which is 0 if the noise is out of sample, otherwise it is the index of the noise. In addition, :noise is the actual type Noise corresponding to the :noiseidx, or the out-of-sample noise as you suggested.

You've probably already dug into the code, but I think you would need to change the type here and then do something clever here.

As for the breakages, I should probably register this package and get some proper version tagging going.

from sddp.jl.

I renamed the keys in my local branch, and it broke my own code! It was a trivial fix, but I can see how it would disrupt other people's code. I believe a way of implementing this change without such issues would be creating a new simulate method, and deprecating the existing ones.

I think having 0 indicate when :noiseidx is sensible, because the output would still always be a (non-negative) Int. It's easy to program around too. I was thinking about :noise being an array with the noise values, but associating it with Noise objects sounds like a better idea.

I could implement the key name changes and SamplingOracles (#80) as well, if you haven't worked on it already. Since you probably already gave it some thought, is there any particular detail you have in mind for them?

from sddp.jl.

I haven't worked on them. They would be a great contribution!

The only thoughts I've had I put down in #80. They seem like a good starting point.

I think once we have sampling oracles, figuring out a new (improved) simulation method will be easier.

I was thinking about :noise being an array with the noise values

This gets a little tricky because I often go

w = [1, 2, 4]
@rhsnoise(sp, i=1:3, x <= w[i])

Should the noise values be 1,2,3 or 1,2,4?

I would suggest making a Pull Request that begins to implement sampling oracles, and we can continue the discussion there. I'm sure we will run into a few problems.

from sddp.jl.