Practical issues in temporal reinforcement learning
	-> Robin
	* Is backgammon a "complex, real world problem"?  Compared to
	  what?
	* What are structural and temporal credit assignment?
	* What does TD stand for?  What does lambda do?
	* What is self-play?  What's an	alternative?
	* How is learning performance measured in this paper?
	* What is feature discovery and why is it important?
	* Why should TDGammon be expected to play better late in the
	  game instead of early in the game?  What does it actually
	  do?

	Generalization in reinforcement learning: Safely approximating
	  the value function
	-> George
	* What is the curse of dimensionality?
	* What are rollouts and why are the safer than one-step
	  backups?
	* What are four classes of behavior of smooth value iteration?
	* Why would you expect a local function approximator to behave
	  better than a global one?
	* Why are multiple rollouts needed for non-deterministic
	  problems?

	Generalization in reinforcement learning: Successful examples
	  using sparse coarse coding
	-> Nishkam
	* How does smooth value iteration differ from Sutton's RL
	  approach?  How do the function approximators differ in the
	  two papers?
	* How are tilings local?  How do they generalize?
	* Why does performance vary with lambda as a u-shaped curve?
	  How about alpha?

	Scaling reinforcement learning toward RoboCup soccer
	-> Vasil
	* Is the Robocup simulator a "complex, real world problem"?
	  How does it compare to a task with real robots?
	* Why do the authors model keepaway as a semi-Markov decision
	  process instead of a standard MDP?
	* What is a representational selection?  Are there examples in
	  which this is automated?
	* How is learning performance measured in the keep-away task?
	* Is self-play used for training in this paper?