PER Beta Annealed Weight

Problem

Compute a single prioritized replay importance-sampling weight with beta annealing.

The schedule is beta = beta0 + (1 - beta0) * (step / total_steps) with step clamped to [0, total_steps], and the weight is (1 / (n * prob))^beta. A non-positive total_steps uses beta = 1.0, and non-positive n or prob returns 0.0.

Hint

Annealing beta from beta0 to 1 gradually removes the importance-sampling bias.

Test cases

case 1

in: [0.2,10,0.4,0,100]

expected: 0.757858283255199

case 2

in: [0.2,10,0.4,100,100]

expected: 0.5

case 3

in: [0.2,10,0.4,50,100]

expected: 0.6155722066724582

case 4

in: [0.2,10,0,50,100]

expected: 0.7071067811865476

case 5

in: [0,10,0.4,50,100]

expected: 0

Starter codePython

def per_beta_annealed_weight(prob, n, beta0, step, total_steps):
    # Your code here
    pass

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