Added getMinAOO which uses Simulated annealing to determine the optim…

[johnWilshire]

…al shift parameters.

You can run the changes after pulling and building with
examples(getMinAOO)

One problem that might come up is that I dont think the values of xshift and yshift are constrained but I havent run into any issues yet.

[calvinkflee]

I tried running the function with a test dataset I have which is a subset of mangroves in Victoria (it has a dimension of 1167x1502, resolution at 30m) and running it at the default maxiter (10,000) took way too long. I've cut that down to 500 to try it out, but it didn't manage to capture a good AOO.
I'm playing with the parscale function since moving the grid 1m at a time when the grid.size is 10,000 (and at 500 iterations I'm not surprised I'm not getting a good minAOO), adding a parscale of grid.size/20.
This led to the whole window going out of bounds, as John suggested might happen, since xshift and yshift is unconstrained. Is there anyway it can be constrained?
At my current parameters, it still took just under 4 minutes to run, and it's still quite a small dataset, so I actually am thinking of even smaller maxiter/larger parscale, but this will not work if the window just loses focus every time.

[johnWilshire]

Would you be able to upload the dataset as part of the package so we can use in the @examples? and so I can have something to test it on.

The SANN method of optim is pretty slow but the function we are trying to optimize is very flat a lot of the time so the much faster gradient methods tend to think that we are in a global min all the time based off local flat information.

In the latest commit I have implemented it as a call to constrOptim with max it = 500 it seems to be running a lot faster.
I am assuming the bounds for each parameter are 0, grid.size?

[calvinkflee]

in the previous line: grid = shift(grid, x = x[1], y = x[2]), doesn't x[2] just make the value default to NA?
Since optim() only takes 1 variable for optimization, doesn't that mean it just searches diagonally (x=y)? Or am I completely misunderstanding what x[2] does...

[johnWilshire]

objective_function takes a vector
run print(o) before the function exits to see, $par is the argmin

[calvinkflee]

Bounds of either (0, grid.size) or (-0.5grid.size, 0.5grid.size) will both be good.

[calvinkflee]

From what I understand of the parscale parameter, doesn't this mean that the adjustments moved will be too big (the whole grid we're working at)?

[calvinkflee]

Thanks John, I can't say that I 100% understand how the barrier function works, but it looks good!

I was wondering if there's a way of forcing o$par to be an integer, rather than numeric? It doesn't make much sense to be varying at those levels when our bounds are at 10,000. Might also give us better results if it's running at larger steps (as opposed to 3d.p.)

I'm also a bit worried about the run-time because the outer.iterations I got here was 3, which adds up to ~10 minutes for a relatively tiny raster. However, the function won't really work if we only run it once, right?