MARWIL Weighted BC

Problem

Compute the MARWIL weighted behavioral cloning loss.

The weight of each sample is exp(beta * advantage), and the loss is the negative weighted mean of the log-probabilities. A zero total weight or empty batch returns 0.0.

Hint

Beta controls how strongly the advantage tilts the imitation weights.

Test cases

case 1

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

expected: 0.75

case 2

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

expected: 0.5596014610110588

case 3

in: [[],[],1]

expected: 0

Starter codePython

import math
def marwil_weighted_loss(log_probs, advantages, beta):
    # Your code here
    pass

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