Target function is evaluated before the optimization is run

[sharpsy]

When defining toy example like this:

import GPyOpt

def f(x):
    print(x)
    return (2*x)**2

bounds = [{'name': 'var_1', 'type': 'continuous', 'domain': (-1, 1)}]

task = GPyOpt.methods.BayesianOptimization(f, bounds)

We can see that the function is already evaluated five times. This function can be arbitrarily expensive to calculate and I believe it is a bug if it is implicitly evaluated five times before the optimization is run.

Using GPyOpt version 1.2.1 with GPy version 1.8.5 on Python 3.5.2

[cookiees0]

I'm not a dev, but there is an option "initial_design_numdata" which is defaulted to 5, and "initial_design_type" which is set to "random", since you need some amount of initial points. Both are easily changed and work well for me.

[sharpsy]

Tnx for clarification. I understand this design choice and I still think that this behavior looks unexpected; by defining the method for the optimization it already evaluates the target function. I accept that some amount of data points are needed in order for the method to work and would like for it to happen when explicitly told to. That way it could also take max_time, max_iter and other optimization parameters into account when the optimization is run that cannot be set with the constructor.

I can try to set initial_design_numdata to 0 to workaround this and increase the max_iter to account for it.

[farquet]

As a user, I also agree that this default value is counter-intuitive and should be set to 0.

[amarazov]

I also think that the default should be 0.