6.py

import numpy as np
from scipy.spatial.distance import pdist
from scipy.linalg import pinv

def rbf_train(K, X, T):
    # input:
    # M x N training samples X
    # 1 x N target outputs T
    #
    # output:
    # K x M prototype vectors mu
    # 1 x 1 scalar Gaussian with sigma
    # K x 1 output weights w

    [M,N] = X.shape

    # select K datapoints at random

    mu = np.random.shuffle(X.transpose())[0:K]

    # compute global distance parameter
    
    dmax = reduce(lambda acc, (mu1, mu2): max(acc, pdist(mu1, mu2)),
                  itertools.combinations(mu, 2))
    
    sigma = dmax/((2*K)**.5)
        
    # compute w
 
    # Add your solution here.
    
    return [mu, sigma, w]