PILCO Software Package V0.9 (2013-07-04)
I. Introduction This software package implements the PILCO RL policy search framework. The learning framework can be applied to MDPs with continuous states and controls/actions and is based on probabilistic modeling of the dynamics and approximate Bayesian inference for policy evaluation and improvement.
II. Quick Start We have already implemented some scenarios that can be found in /scenarios .
If you want to get started immediately, go to /scenarios/cartPole
and execute
cartPole_learn
III. Documentation A detailed documentation can be found in
/doc/doc.pdf
which also includes a description of how to set up your own scenario (there are only a few files that are scenario specific).
IV. Contact If you find bugs, have questions, or want to give us feedback, please send an email to m.deisenroth@imperial.ac.uk
V. References M.P. Deisenroth and C.E. Rasmussen: PILCO: A Data-Efficient and Model-based Approach to Policy Search (ICML 2011) M.P. Deisenroth: Efficient Reinforcement Learning Using Gaussian Processes (KIT Scientific Publishing, 2010)
Marc Deisenroth Andrew McHutchon Joe Hall Carl Edward Rasmussen