Policy Gradient Loss

Problem

Compute the vanilla policy-gradient loss for a batch.

Return the mean of -log_probs[i] * returns[i]. An empty batch gives 0.0.

Hint

Minimizing this loss is gradient ascent on expected return.

case 1

in: [[-0.5,-1],[2,3]]

expected: 2

case 2

in: [[0],[5]]

expected: 0

case 3

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

expected: -2

case 4

in: [[],[]]

expected: 0

Starter codePython

def policy_gradient_loss(log_probs, returns):
    # Your code here
    pass

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