SAC Log Prob Tanh Correction

Problem

Correct a Gaussian log-probability for tanh squashing.

Subtract the sum of log(1 - tanh(u)^2) over the pre-tanh actions, using the stable identity 2*(ln 2 - u - ln(1 + exp(-2u))) for each u.

Hint

The change-of-variables term keeps the squashed density normalized.

case 1

in: [0,[0]]

expected: 0

case 2

in: [0,[1]]

expected: 0.8675616609660544

case 3

in: [-1,[0.5,-0.5]]

expected: -0.51954197216689

case 4

in: [0,[]]

expected: 0

Starter codePython

import math
def sac_tanh_log_prob(base_log_prob, pre_tanh_actions):
    # Your code here
    pass

4 test cases · run your code against them in the editor