GP Kernel

[RodrigoAVargasHdz]

Hi,

I like your code :)

I was wondering if you can implement the use of kernels with individual parameters for each feature. This tends to make the GPs more accurate.

Thanks :)

[ppgaluzio]

Hey, thanks!

I guess I could do it by including an extra parameter at the init method of the class, in which case you could give a kernel object to the optimizer with the parameter of your choice. This way the interface wouldn't be changed and the user has more flexibility in the implementation.

[RodrigoAVargasHdz]

Hi!

Yeah you could do something like this,

        self.GP[i] = GPR(length_scale=np.ones(self.x_dim), 
        				kernel=C(1.0, (1e-3, 1e3)) * Matern(nu=2.5), 
                         n_restarts_optimizer=self.n_rest_opt)

where x_dim is the number of features and C is the constant kernel. May have to declare it,

     from sklearn.gaussian_process.kernels import Matern, ConstantKernel as C 

Thanks!

[RodrigoAVargasHdz]

I had a mistake,

    self.GP[i] = GPR(kernel=C(1.0, (1e-3, 1e3)) * Matern(length_scale=np.ones(self.x_dim), nu=2.5), 
                     n_restarts_optimizer=self.n_rest_opt)

where "length_scale=np.ones(self.x_dim)" indicates that the kernel function has independent length parameters for each individual feature.

[ppgaluzio]

I was thinking more of something like:

def __init__(self, ..., kernel=None):

    if kernel is None:
        self._kernel = Matern(nu) # current default
    else:
        self._kernel = kernel

    self.GP[i] = GPR(kernel=self._kernel)

so that the user can pass the kernel he wants or use the default, in this case, I think the kernel parameter could even be a list of kernel objects, in which case each objective could have a different kernel.

[RodrigoAVargasHdz]

Yeah that could work too!
You will have to import all different types of kernels from the sklearn library.

[RodrigoAVargasHdz]

Hi,
I saw that you updated the coda and include the possibility to use more robust kernels, Thanks :)
I should define the kernel function using the standard notation from Sklearn, correct?

Again, thanks :)

[ppgaluzio]

Hi, yeah, I just didn't have time to test it yet, but it should work without a problem.
You just define the any kernel from sklearn, the argument is passed directly to the GP, so any kernel implemented in sklearn is gonna work.