Created
January 24, 2017 23:59
-
-
Save mrecos/e5d9124656382d9acd2886d2a1b22511 to your computer and use it in GitHub Desktop.
Analytical solution to Kernel Ridge Regression. Process: 1) Simulate N data points, 2) define N x N kernel as desired (RBF here), 3) Perform KRR by regularizing kernel by lambda and solving for 'y', 4) estimate response as y = K %*% \alpha
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
### Simualte some one-dimensional data | |
# Constants | |
a = 50 | |
b = 50 | |
c = 80 | |
N = 10 # low dimensions help to visualize matrix | |
#Limits | |
x_upper <- 100 | |
x_lower <-.01 | |
spacing = (x_upper-x_lower)/(N-1) | |
x <- seq(x_lower,x_upper,by= spacing) | |
# Response | |
y = 0.5*(sin(x-a)/(x-a)) + 0.8*(sin(x-b)/(x-b)) + .3*(sin(x-c)/(x-c)) + rnorm(x,0,0.05) | |
### Prepare the Gaussian RBF kernel | |
N <- length(x) | |
kk <- tcrossprod(x) | |
dd <- diag(kk) | |
ident.N <- diag(rep(1,N)) | |
#RBF Parameters | |
sigma = 0.5 | |
lambda = 0.001 | |
## Gaussian RBF kernel manually | |
myRBF.kernel <- exp(sigma*(-matrix(dd,N,N)-t(matrix(dd,N,N))+2*kk)) | |
### Train KRR parameters from train data | |
alphas <- solve(myRBF.kernel + lambda*ident.N) | |
alphas <- alphas %*% y | |
### above is the same as: | |
## alphas <- solve(myRBF.kernel + lambda*ident.N) %*% y | |
## alphas <- solve(myRBF.kernel + lambda*ident.N, y) | |
### estimate response of training data: | |
yhat <- myRBF.kernel %*% alphas |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment