Universal RL: Applications and Approximations

While the main ideas underlying Universal RL have existed for over a decade now (see [Hutter, 2012] for historical context), practical applications are only just starting to emerge. In particular, the direct approximation introduced by Veness et al. [2010, 2011] was shown empirically to compare favorably to a number of other model-based RL techniques on small, partially observable environments with initially unknown, stochastic dynamics. Since then, a variety of additional techniques have been introduced that allow for the construction of far more sophisticated approximations. This short paper collects together and reviews some of the main ideas that have the potential to lead to larger scale applications.

Authors

• Joel Veness

Keywords

No keywords are indexed for this paper.