pass value_scale to the LearnerND's loss_per_simplex function

[basnijholt]

Currently, the LearnerND cannot access the unscaled value in the loss_per_simplex.

This PR fixes this by passing the scale. This fixes a regression introduced when we added scaling.

Now the following code:

importnumpyasnpimportadaptiveadaptive.notebook_extension()
defisoline_loss_function(y_iso, sigma, priority=1):
fromadaptive.learner.learnerNDimportdefault_lossdefgaussian(x, mu, sigma):
returnnp.exp(-(x-mu) **2/sigma**2)
defloss(simplex, values, value_scale):
distance=np.mean([abs(y_iso*value_scale-y) foryinvalues])
returnpriority*gaussian(distance, 0, sigma) +default_loss(simplex, values, value_scale)
returnlosslevel=0.5loss=isoline_loss_function(y_iso=level, sigma=1, priority=0.3)
deff(xy):
x, y=xyreturnx**2+y**2/2+y**3learner=adaptive.LearnerND(
f,
bounds=[(-1, 1), (-1, 1)],
loss_per_simplex=loss,
)
adaptive.runner.simple(learner, goal=lambdal: l.npoints>1000)
%optsImage [colorbar=True]
learner.plot_isoline(level, tri_alpha=0.3)

produces

instead of

[jbweston]

Seems ok, don't know why tests are failing

[jbweston]

@basnijholt the test failures are genuine.

> loss_curvature = triangle_loss(simplex, values, neighbors, neighbor_values)
E TypeError: triangle_loss() missing 1 required positional argument: 'neighbor_values'

Threre may be others

[basnijholt]

Thanks! I've already pushed a fix.

[basnijholt]

@jbweston merge if you are OK with this 👍