Entropy Bonus Term In Loss

Problem

Combine a policy-gradient term with an entropy bonus.

Return -mean(log_probs[i] * advantages[i]) - coef * mean(entropies). An empty batch returns 0.0.

Hint

The entropy bonus is subtracted so it rewards keeping the policy stochastic.

case 1

in: [[-1,-1],[1,1],[0.5,0.5],0.1]

expected: 0.95

case 2

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

expected: 2

case 3

in: [[],[],[],0.1]

expected: 0

Starter codePython

def entropy_regularized_loss(log_probs, advantages, entropies, coef):
    # Your code here
    pass

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