Prioritized TD Error

Problem

Convert TD errors into non-negative replay priorities.

Each priority is (|delta| + epsilon) ** alpha, so zero errors still receive a small non-zero priority. Return a list of the same length; an empty input returns [].

Hint

High-advantage and high-error transitions get replayed more often.

Test cases

case 1

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

expected: [1.000001,2.000001]

case 2

in: [[0],0.5]

expected: [0.001]

case 3

in: [[],0.5]

expected: []

case 4

in: [[4,9],0.5]

expected: [2.0000002499999843,3.000000166666662]

Starter codePython

def prioritized_td_error(td_errors, alpha, epsilon=1e-6):
    # Your code here
    pass

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