Assignment problem

The assignment problem is a fundamental combinatorial optimization problem. It consists of finding, in a weighted bipartite graph, a matching of a given size, in which the sum of weights of the edges is a minimum.

A common variant consists of finding a maximum-weight matching. This is a specialization of the maximum weight matching problem, in which the input graph is bipartite.

In its most general form, the problem is as follows:

The problem instance has a number of agents and a number of tasks. Any agent can be assigned to perform any task, incurring some cost that may vary depending on the agent-task assignment. It is required to perform as many tasks as possible by assigning at most one agent to each task and at most one task to each agent, in such a way that the total cost of the assignment is minimized.

If the numbers of agents and tasks are equal, then the problem is called balanced assignment. Otherwise, it is called unbalanced assignment.[1]

If the total cost of the assignment for all tasks is equal to the sum of the costs for each agent (or the sum of the costs for each task, which is the same thing in this case), then the problem is called linear assignment.

Commonly, when speaking of the assignment problem without any additional qualification, then the linear balanced assignment problem is meant.

ExamplesEdit

Suppose that a taxi firm has three taxis (the agents) available, and three customers (the tasks) wishing to be picked up as soon as possible. The firm prides itself on speedy pickups, so for each taxi the "cost" of picking up a particular customer will depend on the time taken for the taxi to reach the pickup point. This is a balanced assignment problem. Its solution is whichever combination of taxis and customers results in the least total cost.

Now, suppose that there are four taxis available, but still only three customers. This is an unbalanced assignment problem. One way to solve it is to invent a fourth dummy task, perhaps called "sitting still doing nothing", with a cost of 0 for the taxi assigned to it. This reduces the problem to a balanced assignment problem, which can then be solved in the usual way and still give the best solution to the problem.

Similar adjustments can be done in order to allow more tasks than agents, tasks to which multiple agents must be assigned (for instance, a group of more customers than will fit in one taxi), or maximizing profit rather than minimizing cost.


Formal definitionEdit

The formal definition of the assignment problem (or linear assignment problem) is

Given two sets, A and T, of equal size, together with a weight function C : A × TR. Find a bijection f : AT such that the cost function:
 

is minimized.

Usually the weight function is viewed as a square real-valued matrix C, so that the cost function is written down as:

 

The problem is "linear" because the cost function to be optimized as well as all the constraints contain only linear terms.

AlgorithmsEdit

A naive solution for the assignment problem is to check all the assignments and calculate the cost of each one. This may be very inefficient since, with n agents and n tasks, there are n! (factorial of n) different assignments. Fortunately, there are many algorithms for solving the problem in time polynomial in n.

The assignment problem is a special case of the transportation problem, which is a special case of the minimum cost flow problem, which in turn is a special case of a linear program. While it is possible to solve any of these problems using the simplex algorithm, each specialization has more efficient algorithms designed to take advantage of its special structure.

Balanced assignmentEdit

In the balanced assignment problem, both parts of the bipartite graph have the same number of vertices, denoted by n.

One of the first polynomial-time algorithms for balanced assignment was the Hungarian algorithm. It is a global algorithm - it is based on improving a matching along augmenting paths (alternating paths between unmatched vertices). Its run-time complexity, when using Fibonacci heaps, is  .[2] This is currently the fastest run-time of a strongly polynomial algorithm for this problem. If all weights are integers, then the run-time can be improved to  , but the resulting algorithm is only weakly-polynomial.[3] If the weights are integers, and all weights are at most C (where C>1 is some integer), then the problem can be solved in  weakly-polynomial time in a method called weight scaling.[4][5][6]

In addition to the global methods, there are local methods which are based on finding local updates (rather than full augmenting paths). These methods have worse asymptotic runtime guarantees, but they often work better in practice. These algorithms are called auction algorithms, push-relabel algorithms, or preflow-push algorithms. Some of these algorithms were shown to be equivalent.[7]

Some of the local methods assume that the graph admits a perfect matching; if this is not the case, then some of these methods might run forever.[1]:3 A simple technical way to solve this problem is to extend the input graph to a complete bipartite graph, by adding artificial edges with very large weights. These weights should exceed the weights of all existing matchings, to prevent appearance of artificial edges in the possible solution.

As shown by Mulmuley, Vazirani and Vazirani,[8] the problem of minimum weight perfect matching is converted to finding minors in the adjacency matrix of a graph. Using the isolation lemma, a minimum weight perfect matching in a graph can be found with probability at least ½. For a graph with n vertices, it requires   time.

Unbalanced assignmentEdit

In the unbalanced assignment problem, the larger part of the bipartite graph has n vertices and the smaller part has r<n vertices. There is also a constant s which is at most the maximum cardinality of a matching in the graph. The goal is to find a minimum-cost matching of size exactly s. The most common case is the case in which the graph admits a one-sided-perfect matching (i.e., a matching of size r), and s=r.

Unbalanced assignment can be reduced to a balanced assignment. The naive reduction is to add   new vertices to the smaller part and connect them to the larger part using edges of cost 0. However, this requires   new edges. A more efficient reduction is called the doubling technique. Here, a new graph G' is built from two copies of the original graph G: a forward copy Gf and a backward copy Gb. The backward copy is "flipped", so that, in each side of G', there are now n+r vertices. Between the copies, we need to add two kinds of linking edges:[1]:4-6

  • Large-to-large: from each vertex in the larger part of Gf, add a zero-cost edge to the corresponding vertex in Gb.
  • Small-to-small: if the original graph does not have a one-sided-perfect matching, then from each vertex in the smaller part of Gf, add a very-high-cost edge to the corresponding vertex in Gb.

All in all, at most   new edges are required. The resulting graph always has a perfect matching of size  . A minimum-cost perfect matching in this graph must consist of minimum-cost maximum-cardinality matchings in Gf and Gb. The main problem with this doubling technique is that there is no speed gain when  .

Instead of using reduction, the unbalanced assignment problem can be solved by directly generalizing existing algorithms for balanced assignment. The Hungarian algorithm can be generalized to solve the problem in   strongly-polynomial time. In particular, if s=r then the runtime is  . If the weights are integers, then Thorup's method can be used to get a runtime of  . [1]:6

Solution by linear programmingEdit

The assignment problem can be solved by presenting it as a linear program. For convenience we will present the maximization problem. Each edge (i,j), where i is in A and j is in T, has a weight  . For each edge (i,j) we have a variable  . The variable is 1 if the edge is contained in the matching and 0 otherwise, so we set the domain constraints:    

The total weight of the matching is:  . The goal is to find a maximum-weight perfect matching.

To guarantee that the variables indeed represent a perfect matching, we add constraints saying that each vertex is adjacent to exactly one edge in the matching, i.e,  .

All in all we have the following LP:

 
 
 
 
This is an integer linear program. However, we can solve it without the integrality constraints (i.e., drop the last constraint), using standard methods for solving continuous linear programs. While this formulation allows also fractional variable values, in this special case, the LP always has an optimal solution where the variables take integer values. This is because the constraint matrix of the fractional LP is totally unimodular - it satisfies the four conditions of Hoffman and Gale.

This can also be proved directly.[9]:31-37 Let x be an optimal solution of the fractional LP, w(x) be its total weight, and k(x) be the number of non-integral variables. If k(x)=0 we are done. Otherwise, there is a fractional variable, say  . Because the sum of variables adjacent to j2 is 1, which in an integer, there must be another variable adjacent to j2 with a fractional value, say  . By similar considerations on i3, there must be another variable adjacent to i3 with a fractional value, say  . By similar considerations we move from one vertex to another, collecting edges with fractional values. Since the graph is finite, at some point we must have a cycle. Without loss of generality we can assume that the cycle ends at vertex i1, so the last fractional variable in the cycle is  . So the number of edges in the cycle is 2m - it must be even since the graph is bipartite.

Suppose we add a certain constant e to all even variables in the cycle, and remove the same constant e from all odd variables in the cycle. For any such e, the sum of variables near each vertex remains the same (1), so the vertex constraints are still satisfied. Moreover, if e is sufficiently small, all variables remain between 0 and 1, so the domain constraints are still satisfied too. It is easy to find a largest e that maintains the domain constraints: it is either the smallest difference between an odd variable and 0, or the smallest difference between an even variable and 1. Now, we have one less fractional variable, so k(x) decreases by 1. The objective value remains the same, since otherwise we could increase it by selecting e to be positive or negative, in contradiction to the assumption that it is maximal.

By repeating the cycle-removal process we arrive, after at most n steps, at a solution in which all variables are integral.

See alsoEdit

References and further readingEdit

  1. ^ a b c d Lyle Ramshaw, Robert E. Tarjan (2012). "On minimum-cost assignments in unbalanced bipartite graphs" (PDF). HP research labs.
  2. ^ Fredman, Michael L.; Tarjan, Robert Endre (1987-07-01). "Fibonacci Heaps and Their Uses in Improved Network Optimization Algorithms". J. ACM. 34 (3): 596–615. doi:10.1145/28869.28874. ISSN 0004-5411.
  3. ^ Thorup, Mikkel (2004-11-01). "Integer priority queues with decrease key in constant time and the single source shortest paths problem". Journal of Computer and System Sciences. Special Issue on STOC 2003. 69 (3): 330–353. doi:10.1016/j.jcss.2004.04.003. ISSN 0022-0000.
  4. ^ Gabow, H.; Tarjan, R. (1989-10-01). "Faster Scaling Algorithms for Network Problems". SIAM Journal on Computing. 18 (5): 1013–1036. doi:10.1137/0218069. ISSN 0097-5397.
  5. ^ Goldberg, A.; Kennedy, R. (1997-11-01). "Global Price Updates Help". SIAM Journal on Discrete Mathematics. 10 (4): 551–572. doi:10.1137/S0895480194281185. ISSN 0895-4801.
  6. ^ Orlin, James B.; Ahuja, Ravindra K. (1992-02-01). "New scaling algorithms for the assignment and minimum mean cycle problems". Mathematical Programming. 54 (1–3): 41–56. doi:10.1007/BF01586040. ISSN 0025-5610.
  7. ^ Vargas, Marcos C.; Valencia, Carlos E.; Perez, Sergio L.; Alfaro, Carlos A. (2018-10-08). "The equivalence between two classic algorithms for the assignment problem". arXiv:1810.03562v1. Bibcode:2018arXiv181003562A. Cite journal requires |journal= (help)
  8. ^ Mulmuley, Ketan; Vazirani, Umesh; Vazirani, Vijay (1987). "Matching is as easy as matrix inversion". Combinatorica. 7 (1): 105–113. doi:10.1007/BF02579206.
  9. ^ Gärtner, Bernd; Matoušek, Jiří (2006). Understanding and Using Linear Programming. Berlin: Springer. ISBN 3-540-30697-8.