In my post at the RichRelevance Engineering Blog, I describe how one can use Gaussian Process Regression for online learning. Excerpt:
Consider a situation where you have a dial to tweak, and this dial setting may influence a reward of some kind. For example, the dial may be a weight used in a personalization algorithm, and the reward may be clickthrough or revenue. The problem is, we don’t know beforehand how the dial affects the reward, and the reward behavior may be noisy. How then can we choose a dial setting that maximizes the reward?
A handy way to approach this problem is to model the unknown reward function as an instance of a Gaussian Process – this method is called kriging or Gaussian Process Regression (GPR)…
Read more here.