Least Squares Policy Iteration
 Lagoudakis and Parr

What is on-policy vs. off-policy?
Why is a state-value function of no use to control?
What did the authors have to do to extend LSTD to LSQ?
What does P^pi Phi represent?  Is it constructable in a continuous
	state space?  How about hat{P^pi Phi}?
How is hat{A} used differently given changes in the policy?
Is this method memory based?  What experience needs to be stored?
Problems solved: linear chain, inverted pendulum, bicycle task.
	Would a direct comparison with other methods be meaningful?
	Is number of episodes to master the task a useful comparison
	metric?
How was training data collected?  Is this realistic?
What is the Ormoneit and Sen paper about?