Softmax Policy Probabilities

Problem

Convert action preferences into a probability distribution with a temperature parameter.

Return softmax(logits / temperature) computed with the max-subtraction trick so large logits do not overflow. The temperature is positive and logits is non-empty.

Hint

Subtracting the maximum leaves the ratio unchanged but keeps exp() in range.

Test cases

case 1

in: [[0,0],1]

expected: [0.5,0.5]

case 2

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

expected: [0.09003057317038046,0.24472847105479764,0.6652409557748218]

case 3

in: [[1000,1000],1]

expected: [0.5,0.5]

case 4

in: [[2,1,0],0.5]

expected: [0.8668133321973349,0.11731042782619838,0.015876239976466765]

case 5

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

expected: [0.3333333333333333,0.3333333333333333,0.3333333333333333]

Starter codePython

import math
def softmax_probs(logits, temperature):
    # Your code here
    pass

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