Suggesting Values

[kjans123]

Steps to reproduce

loc = ng.p.Scalar(lower=-5,upper=5)
scale = ng.p.Scalar(lower=0, upper=5)
s = ng.p.Scalar(lower=0, upper=10)
k = ng.p.Choice(list(range(2,6)))
w = ng.p.Array(shape=(self.times.shape[0],)).set_bounds(-10,10)
instru = ng.p.Instrumentation(loc=loc,
                          scale = scale,
                          s=s,
                          k=k,
                          w = w)
optimizer = ng.optimizers.DE(parametrization=instru,
                                      budget=budget)
optimizer.suggest((),{'k':3,'loc':-2,'s':2,'scale':2,'w':np.ones(self.times.shape[0])})

Observed Results

ValueError: Tuple value must be a tuple of size 0, got: ((), {'k': 3, 'loc': -2, 's': 2, 'scale': 2, 'w': array([1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
       1., 1., 1., 1., 1., 1.])}).
Current value: ()

Expected Results

For initial values to be set in an optimizer run

Has anyone had success using the suggest method? If so, would you mind copying/pasting working code? I've been trying different forms of the example in the documentation, but cannot seem to get it to work.

[jrapin]

Hi,
Can you point to the part of the documentation you have been looking at? So that we can clarify it.

Basically, suggest should be called the same way as the function to optimize, in your case, given you are using an Instrumentation, I guess it should be:
optimizer.suggest(k=3, loc=-2, s=2, scale=2, w=mp.ones(self.times.shape[0]))

Another option, which can work for all but the Choice parameter would be to use the init option of Array and Scalar (eg: loc = ng.p.Scalar(init=-2, lower=-5, upper=5))

[kjans123]

Here is the part I was looking at:
https://facebookresearch.github.io/nevergrad/optimization.html?highlight=suggest#telling-non-asked-points-or-suggesting-points

And you're suggested edit did the trick. Thanks!