(i)   score identity  grad log pi(a) = e_a - pi : verified for all actions
(ii)  E_pi[score] = [ 0.  0.  0. -0. -0.]
(iii) grad J (finite diff) : [ 0.187313 -0.009418  0.074081  0.004758 -0.256735]
      pi * (R - J)         : [ 0.187313 -0.009418  0.074081  0.004758 -0.256735]
      sampled estimate     : [ 0.187559 -0.009321  0.073992  0.004821 -0.257051]
      standard error       : [5.16e-04 1.44e-04 3.01e-04 4.50e-05 4.21e-04]
