Link Search Menu Expand Document
POLAR

There are several hyperparameters used in POLAR. Including the following major hyperparameters:

Contents
  1. Lengthscales
  2. GP Noise Variance
  3. GP Feedback Noise
  4. Upper Confidence Bound for Region of Avoidance

In this tutorial we discuss methods of tuning these hyperparameters.

Lengthscales

The following is an example of three different lengthscales for the 1D Function example in the POLAR toolbox:

Small Lengthscale Medium Lengthscale Large Lengthscale

Within the 1D Function example folder, there is a script called how_to_tune_lengthscale.m that demonstrates the effect of the lengthscale on the learning performance: Learning Performance

As shown by the figure, too small of a lengthscale results in a slower learning rate, while too large of a lengthscale results in overly-confident assumptions about the underlying optimal action.

GP Noise Variance

GP Feedback Noise

Upper Confidence Bound for Region of Avoidance

This setting is stored in the learning algorithm as settings.roa.lambda and appears in the equation used to estimate the region of avoidance.