Prediction for the last point incorrect

[dcferre]

Hello,

I am finding a recurrent issue with the predictions of the optimizer for the last point that was added to the dataset.

I am using the following chunk of code to initialize a BayesianOptimization instance:

Bounded region of parameter space

pbounds = {'x': (-20, 20)}

Bayesian optimization with upper confidence bound acquisition function

ucb = acquisition.UpperConfidenceBound(kappa=2.5)
bo = BayesianOptimization(
f=None,
acquisition_function = ucb,
pbounds=pbounds,
verbose=2,
random_state=2803,
)

Then, I declare the following function that plays the role of an "experiment" and one to add new runs to the dataset:

Exemple of "true" function

def experiment(x):
y = (1/2)np.sin(x/4) - np.cos(x/6) + 2np.sin(x/2)
return y

Add new runs to the optimizer

def next_experiment(bo, ne=1, verbose=1):

for i in range(ne):
    next_run = bo.suggest()
    xi = next_run['x']
    yi = experiment(xi)
    bo.register(params=next_run, target=yi)
    if verbose > 0:
        print(f'x = {xi:6.2f} : y = {yi:6.2f}')
    
return bo

For example, I can add three runs and I get the following output:

bo = next_experiment(bo, ne=3)

x = -4.85 : y = -2.47
x = -17.10 : y = -0.12
x = -17.08 : y = -0.14

However, when I use the predict method of the _gp model, I am not getting back the correct value for the last entry:

xi = bo.space.params.flatten()
yi = bo._gp.predict(xi.reshape(-1,1))
for x, y in zip(xi,yi):
print(f'x = {x:6.2f} : yp = {y:6.2f}')

x = -4.85 : yp = -2.47
x = -17.10 : yp = -0.12
x = -17.08 : yp = -0.87

But, if I add two new runs and I repeat the calculation of the predicted values, then yes, I get the correct value for the third entry (x = -17.08, yp = -0.14), but again, I am getting a wrong value for the last entry (x = 20.00), -0.06 instead of -0.58.

bo = next_experiment(bo, ne=2)

x = -19.92 : y = 2.49
x = 20.00 : y = -0.58

xi = bo.space.params.flatten()
yi = bo._gp.predict(xi.reshape(-1,1))
for x, y in zip(xi,yi):
print(f'x = {x:6.2f} : yp = {y:6.2f}')

x = -4.85 : yp = -2.47
x = -17.10 : yp = -0.12
x = -17.08 : yp = -0.14
x = -19.92 : yp = 2.49
x = 20.00 : yp = -0.06

Am I missing something? Is this the correct behaviour?

Thanks!!

[till-m]

Hey @dcferre,

see #513.

I'm realizing that while I added a note to this effect to the .maximize function docs (link) I never did the same for .probe. Would that have helped clear the confusion in your case?

[dcferre]

Hi @till-m,

Yes, it helps. Thanks a lot for your fast reply.

And sorry for the mess of the layout of my first message, rookie using github.

[till-m]

No problem! Let me know if you run into any other problem :)