weighted.py

# -*- python -*-

'''

SYNOPSIS

    >>> import weighted

    >>> choices = { 'heads': 2, 'tails': 1 }
    >>> for i in xrange(100):
    ...    print weighted.weighted(choices)

    This example would print a hundred coin flips, but the heads
    will appear twice as often as tails, statistically speaking.

AUTHOR

    Ed Halley ([email protected]) 13 December 2007

'''

import random

def pare(choices):
    '''Given a dict of key:weight pairs, returns the total of all weights.'''
    return sum(choices.values())

def weighted(choices, total=0):
    '''Given a dict of key:weight pairs, chooses a key at random.

    The dict values are non-negative numerical weights.  Keys with higher
    values are chosen more often than keys with lower values.

    If the caller knows the total of all weights, it can be given to
    avoid recalculating it internally on each call.  If the given total
    is not accurate, a key may be chosen with a poorly-shaped
    distribution.
    '''
    if not total:
        total = pare(choices)
    mark = random.random()*total
    keys = choices.keys()
    for i in xrange(len(keys)):
        span = choices[keys[i]]
        if span > mark:
            return keys[i]
        mark -= span
    # should not reach here if total is accurate
    return random.choice(keys)

if __name__ == '__main__':
    print 'Testing weighted random distribution...'
    choices = { 'ten': 10,
                'eight': 8,
                'seven': 7,
                'six': 6,
                'four': 4,
                'one': 1 }
    tally = { }
    reps = 10000000
    total = sum(choices.values())
    for i in xrange(reps):
        if 0 == i % 500000: print reps-i, '\r',
        x = weighted(choices, total)
        try: tally[x] += 1
        except: tally[x] = 1
    print reps, 'reps', '==', sum(tally.values()), 'picks'
    print "%s\t%9s %9s %s" % ('key:','picks:',
                               'fair:','bias (ideal 100%):')
    for key in choices:
        expected = choices[key]*reps/total
        print "%s\t%9d %9d (%g%%)" % (key, tally[key],
                                      expected, tally[key]*100./expected)