PSR paper

Designed for problems with reset.  Would this also make sense for
episodic problems more generally?  (That is, the learner can't force a
reset, but the environment occassionally returns to an initial state.)

There's something wrong with Lemma 1.  What are "the linear
independent columns of Z"?  Is it a maximal such set?  What is U in
the lemma (are we assuming the environment can be represented by a
POMDP?)?

We can't quite use rank(i) = rank(i+1) as a stopping criterion in
practice, since there is noise that will make the rank appear larger
than it is.  We probably want to do some sort of scree test to see if
the smallest dimensions are actually significant.  (I suspect this is
what Jaeger referred to in his paper.)

I'm surprised they couldn't prove their 2 dimensional version of my 1
dimensional proof...

IO-OOM paper

What's a learning algorithm for IO-OOMs?  Would it imply a learning
algorithm for PSRs?