Top

munkres module

Introduction

The Munkres module provides an implementation of the Munkres algorithm (also called the Hungarian algorithm or the Kuhn-Munkres algorithm), useful for solving the Assignment Problem.

Assignment Problem

Let C be an n by n matrix representing the costs of each of n workers to perform any of n jobs. The assignment problem is to assign jobs to workers in a way that minimizes the total cost. Since each worker can perform only one job and each job can be assigned to only one worker the assignments represent an independent set of the matrix C.

One way to generate the optimal set is to create all permutations of the indexes necessary to traverse the matrix so that no row and column are used more than once. For instance, given this matrix (expressed in Python):

matrix = [[5, 9, 1],
          [10, 3, 2],
          [8, 7, 4]]

You could use this code to generate the traversal indexes:

def permute(a, results):
    if len(a) == 1:
        results.insert(len(results), a)

    else:
        for i in range(0, len(a)):
            element = a[i]
            a_copy = [a[j] for j in range(0, len(a)) if j != i]
            subresults = []
            permute(a_copy, subresults)
            for subresult in subresults:
                result = [element] + subresult
                results.insert(len(results), result)

results = []
permute(range(len(matrix)), results) # [0, 1, 2] for a 3x3 matrix

After the call to permute(), the results matrix would look like this:

[[0, 1, 2],
 [0, 2, 1],
 [1, 0, 2],
 [1, 2, 0],
 [2, 0, 1],
 [2, 1, 0]]

You could then use that index matrix to loop over the original cost matrix and calculate the smallest cost of the combinations:

minval = sys.maxsize
for indexes in results:
    cost = 0
    for row, col in enumerate(indexes):
        cost += matrix[row][col]
    minval = min(cost, minval)

print minval

While this approach works fine for small matrices, it does not scale. It executes in O(n!) time: Calculating the permutations for an n\ x\ n matrix requires n! operations. For a 12x12 matrix, that's 479,001,600 traversals. Even if you could manage to perform each traversal in just one millisecond, it would still take more than 133 hours to perform the entire traversal. A 20x20 matrix would take 2,432,902,008,176,640,000 operations. At an optimistic millisecond per operation, that's more than 77 million years.

The Munkres algorithm runs in O(n\ ^3) time, rather than O(n!). This package provides an implementation of that algorithm.

This version is based on http://csclab.murraystate.edu/~bob.pilgrim/445/munkres.html

This version was written for Python by Brian Clapper from the algorithm at the above web site. (The Algorithm:Munkres Perl version, in CPAN, was clearly adapted from the same web site.)

Usage

Construct a Munkres object:

from munkres import Munkres

m = Munkres()

Then use it to compute the lowest cost assignment from a cost matrix. Here's a sample program:

from munkres import Munkres, print_matrix

matrix = [[5, 9, 1],
          [10, 3, 2],
          [8, 7, 4]]
m = Munkres()
indexes = m.compute(matrix)
print_matrix(matrix, msg='Lowest cost through this matrix:')
total = 0
for row, column in indexes:
    value = matrix[row][column]
    total += value
    print '(%d, %d) -> %d' % (row, column, value)
print 'total cost: %d' % total

Running that program produces:

Lowest cost through this matrix:
[5, 9, 1]
[10, 3, 2]
[8, 7, 4]
(0, 0) -> 5
(1, 1) -> 3
(2, 2) -> 4
total cost=12

The instantiated Munkres object can be used multiple times on different matrices.

Non-square Cost Matrices

The Munkres algorithm assumes that the cost matrix is square. However, it's possible to use a rectangular matrix if you first pad it with 0 values to make it square. This module automatically pads rectangular cost matrices to make them square.

Notes:

  • The module operates on a copy of the caller's matrix, so any padding will not be seen by the caller.
  • The cost matrix must be rectangular or square. An irregular matrix will not work.

Calculating Profit, Rather than Cost

The cost matrix is just that: A cost matrix. The Munkres algorithm finds the combination of elements (one from each row and column) that results in the smallest cost. It's also possible to use the algorithm to maximize profit. To do that, however, you have to convert your profit matrix to a cost matrix. The simplest way to do that is to subtract all elements from a large value. For example:

from munkres import Munkres, print_matrix

matrix = [[5, 9, 1],
          [10, 3, 2],
          [8, 7, 4]]
cost_matrix = []
for row in matrix:
    cost_row = []
    for col in row:
        cost_row += [sys.maxsize - col]
    cost_matrix += [cost_row]

m = Munkres()
indexes = m.compute(cost_matrix)
print_matrix(matrix, msg='Highest profit through this matrix:')
total = 0
for row, column in indexes:
    value = matrix[row][column]
    total += value
    print '(%d, %d) -> %d' % (row, column, value)

print 'total profit=%d' % total

Running that program produces:

Highest profit through this matrix:
[5, 9, 1]
[10, 3, 2]
[8, 7, 4]
(0, 1) -> 9
(1, 0) -> 10
(2, 2) -> 4
total profit=23

The munkres module provides a convenience method for creating a cost matrix from a profit matrix. Since it doesn't know whether the matrix contains floating point numbers, decimals, or integers, you have to provide the conversion function; but the convenience method takes care of the actual creation of the matrix:

import munkres

cost_matrix = munkres.make_cost_matrix(matrix,
                                       lambda cost: sys.maxsize - cost)

So, the above profit-calculation program can be recast as:

from munkres import Munkres, print_matrix, make_cost_matrix

matrix = [[5, 9, 1],
          [10, 3, 2],
          [8, 7, 4]]
cost_matrix = make_cost_matrix(matrix, lambda cost: sys.maxsize - cost)
m = Munkres()
indexes = m.compute(cost_matrix)
print_matrix(matrix, msg='Lowest cost through this matrix:')
total = 0
for row, column in indexes:
    value = matrix[row][column]
    total += value
    print '(%d, %d) -> %d' % (row, column, value)
print 'total profit=%d' % total

Disallowed Assignments

You can also mark assignments in your cost or profit matrix as disallowed. Simply use the munkres.DISALLOWED constant.

from munkres import Munkres, print_matrix, make_cost_matrix, DISALLOWED

matrix = [[5, 9, DISALLOWED],
          [10, DISALLOWED, 2],
          [8, 7, 4]]
cost_matrix = make_cost_matrix(matrix, lambda cost: (sys.maxsize - cost) if
                                      (cost != DISALLOWED) else DISALLOWED)
m = Munkres()
indexes = m.compute(cost_matrix)
print_matrix(matrix, msg='Lowest cost through this matrix:')
total = 0
for row, column in indexes:
    value = matrix[row][column]
    total += value
    print '(%d, %d) -> %d' % (row, column, value)
print 'total profit=%d' % total

Running this program produces:

Lowest cost through this matrix:
[ 5,  9,  D]
[10,  D,  2]
[ 8,  7,  4]
(0, 1) -> 9
(1, 0) -> 10
(2, 2) -> 4
total profit=23

References

  1. http://www.public.iastate.edu/~ddoty/HungarianAlgorithm.html

  2. Harold W. Kuhn. The Hungarian Method for the assignment problem. Naval Research Logistics Quarterly, 2:83-97, 1955.

  3. Harold W. Kuhn. Variants of the Hungarian method for assignment problems. Naval Research Logistics Quarterly, 3: 253-258, 1956.

  4. Munkres, J. Algorithms for the Assignment and Transportation Problems. Journal of the Society of Industrial and Applied Mathematics, 5(1):32-38, March, 1957.

  5. http://en.wikipedia.org/wiki/Hungarian_algorithm

Copyright and License

Copyright 2008-2016 Brian M. Clapper

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Module variables

var DISALLOWED

Functions

def make_cost_matrix(

profit_matrix, inversion_function)

Create a cost matrix from a profit matrix by calling 'inversion_function' to invert each value. The inversion function must take one numeric argument (of any type) and return another numeric argument which is presumed to be the cost inverse of the original profit.

This is a static method. Call it like this:

.. python:

cost_matrix = Munkres.make_cost_matrix(matrix, inversion_func)

For example:

.. python:

cost_matrix = Munkres.make_cost_matrix(matrix, lambda x : sys.maxsize - x)

:Parameters: profit_matrix : list of lists The matrix to convert from a profit to a cost matrix

inversion_function : function
    The function to use to invert each entry in the profit matrix

:rtype: list of lists :return: The converted matrix

Classes

class Munkres

Calculate the Munkres solution to the classical assignment problem. See the module documentation for usage.

Ancestors (in MRO)

Static methods

def make_cost_matrix(

profit_matrix, inversion_function)

DEPRECATED

Please use the module function make_cost_matrix().

Instance variables

var C

var Z0_c

var Z0_r

var col_covered

var marked

var n

var path

var row_covered

Methods

def __init__(

self)

Create a new instance

def compute(

self, cost_matrix)

Compute the indexes for the lowest-cost pairings between rows and columns in the database. Returns a list of (row, column) tuples that can be used to traverse the matrix.

:Parameters: cost_matrix : list of lists The cost matrix. If this cost matrix is not square, it will be padded with zeros, via a call to pad_matrix(). (This method does not modify the caller's matrix. It operates on a copy of the matrix.)

    **WARNING**: This code handles square and rectangular
    matrices. It does *not* handle irregular matrices.

:rtype: list :return: A list of (row, column) tuples that describe the lowest cost path through the matrix

def pad_matrix(

self, matrix, pad_value=0)

Pad a possibly non-square matrix to make it square.

:Parameters: matrix : list of lists matrix to pad

pad_value : int
    value to use to pad the matrix

:rtype: list of lists :return: a new, possibly padded, matrix