|
Reducing Model Bias in Reinforcement Learning
Marc Deisenroth and Carl Edward Rasmussen
In: Learning and Planning from Batch Time Series Data, 11 Dec 2010, Whistler, BC, Canada.
AbstractModel bias is one of the main reasons why reinforcement learning
(RL) algorithms often need so many trials to successfully learn a
task. Model bias has been known for decades, but no general solution
to this problem has yet been proposed. We shed some light on the
challenges of learning models from data and propose learning
probabilistic models to reduce model bias by faitfully incorporating
the model's uncertainty into planning and policy learning [Edit]
|
