Part A: gradient estimator at theta = 0
exact gradient: [-0.1  0.   0.1]
  b=0.0: mean=[-0.10075  0.00039  0.10036]  total variance=0.3127
  b=0.5: mean=[-9.953e-02 -6.000e-05  9.959e-02]  total variance=0.1468
  b=0.8: mean=[-0.10064  0.00013  0.10051]  total variance=0.2072
Part B: probability of the best arm (mean over 60 seeds)
  step    1: 0.350  (min over seeds 0.319)
  step   50: 0.888  (min over seeds 0.814)
  step  100: 0.955  (min over seeds 0.939)
  step  200: 0.979  (min over seeds 0.967)
  step  400: 0.991  (min over seeds 0.986)
