IQL Expectile Loss

Problem

Compute the IQL expectile regression loss.

For each sample diff = q - target and the weight is tau when diff >= 0, otherwise 1 - tau; the loss is the mean of weight * diff^2. An empty batch returns 0.0.

Hint

The expectile controls how pessimistic the value estimate is on the offline data.

Test cases

case 1

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

expected: 0.5

case 2

in: [[2,1],[1,2],0.7]

expected: 0.5

case 3

in: [[3],[3],0.9]

expected: 0

case 4

in: [[],[],0.5]

expected: 0

Starter codePython

def iql_expectile_loss(q_values, targets, tau):
    # Your code here
    pass

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