SAC Alpha Loss

Problem

Compute the SAC temperature (alpha) loss for a batch of action log-probabilities.

J(alpha) = -alpha * (mean(log_probs) + target_entropy). An empty batch returns 0.0.

Hint

The temperature is adjusted so the mean entropy matches the target entropy.

case 1

in: [0.2,[-0.5,-1],-1]

expected: 0.35000000000000003

case 2

in: [0.2,[-0.5,-1],-2]

expected: 0.55

case 3

in: [1,[-0.3],-1]

expected: 1.3

case 4

in: [0,[-0.5,-1],-2]

expected: 0

case 5

in: [0.5,[],-1]

expected: 0

Starter codePython

def sac_alpha_loss(alpha, log_probs, target_entropy):
    # Your code here
    pass

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