Ex 11.3: tokens per sample of the Monte Carlo step-value estimator, n + N n(n-1)/2 + N n, for n = 6
  N=1: 27  N=2: 48  N=4: 90  N=16: 342

Ex 11.5: probability that a group of G responses contains a success, for P(success) = 0.4^6 and 0.25^6
  p0 = 0.4: P(success) = 0.00410   G=4: 0.0163   G=32: 0.1231   G=256: 0.6503
  p0 = 0.25: P(success) = 0.00024   G=4: 0.0010   G=32: 0.0078   G=256: 0.0606

Ex 11.6: step labels can disagree with the outcome.  Two steps; step 1 in {A, B}, step 2 in {X, Y}.
  Outcome reward R = 1 only for (B, X).  A reviewer labels the steps (A: ok, B: not ok), (X: ok, Y: not ok).
  (A, X): outcome 0   sum of step labels 2
  (A, Y): outcome 0   sum of step labels 1
  (B, X): outcome 1   sum of step labels 1
  (B, Y): outcome 0   sum of step labels 0
  maximizer of the outcome: ('B', 'X') (R = 1);  maximizer of the label sum: ('A', 'X') (R = 0)

Ex 11.8: the base rate of a step classifier.  Under a uniform policy over 4 actions, P(step correct) = 1/4, so the classifier
  that always answers 'incorrect' has accuracy 0.75 on step labels; a trained model with accuracy below that is worse than no model.
