/PILCO-disentangling-uncertainty

Disentangling Sources of Uncertainty for Active Exploration (Reinforcement Learning)

Primary LanguageMATLABOtherNOASSERTION

PILCO - disentangling sources of uncertainty

This repository contains the orgininal PILCO MATLAB code (see pilco_readme.txt and license) with added functionality that disentangles the sources of uncertainty in the cost function for a given policy by Monte Carlo sampling.

Uncertainty analysis implemented for the following scenarios:

  • Pendulum
  • Cart-Pole
  • Pendubot
  • Cart-Double-Pendulum

Run

Scenarios can be run from from the scenarios folder:

  • pendulum_learn.m
  • cartPole_learn.m
  • pendubot_learn.m
  • cartDouble_learn.m

Set the number of Monte Carlo samples in the scenario settings files.

Plot

Plots of the Monte Carlo samples and uncertainty decomposition can be made for each environment from the scripts in the uncertainty_plots folder.

Output

example plots for the pendubot scenario:

example plots for the cartDoublePendulum scenario: