explain the hungarian method for solving assignment problem can also be used to solve

Hungarian Method

The Hungarian method is a computational optimization technique that addresses the assignment problem in polynomial time and foreshadows following primal-dual alternatives. In 1955, Harold Kuhn used the term “Hungarian method” to honour two Hungarian mathematicians, Dénes Kőnig and Jenő Egerváry. Let’s go through the steps of the Hungarian method with the help of a solved example.

Hungarian Method to Solve Assignment Problems

The Hungarian method is a simple way to solve assignment problems. Let us first discuss the assignment problems before moving on to learning the Hungarian method.

What is an Assignment Problem?

A transportation problem is a type of assignment problem. The goal is to allocate an equal amount of resources to the same number of activities. As a result, the overall cost of allocation is minimised or the total profit is maximised.

Because available resources such as workers, machines, and other resources have varying degrees of efficiency for executing different activities, and hence the cost, profit, or loss of conducting such activities varies.

Assume we have ‘n’ jobs to do on ‘m’ machines (i.e., one job to one machine). Our goal is to assign jobs to machines for the least amount of money possible (or maximum profit). Based on the notion that each machine can accomplish each task, but at variable levels of efficiency.

Hungarian Method Steps

Check to see if the number of rows and columns are equal; if they are, the assignment problem is considered to be balanced. Then go to step 1. If it is not balanced, it should be balanced before the algorithm is applied.

Step 1 – In the given cost matrix, subtract the least cost element of each row from all the entries in that row. Make sure that each row has at least one zero.

Step 2 – In the resultant cost matrix produced in step 1, subtract the least cost element in each column from all the components in that column, ensuring that each column contains at least one zero.

Step 3 – Assign zeros

• Analyse the rows one by one until you find a row with precisely one unmarked zero. Encircle this lonely unmarked zero and assign it a task. All other zeros in the column of this circular zero should be crossed out because they will not be used in any future assignments. Continue in this manner until you’ve gone through all of the rows.
• Examine the columns one by one until you find one with precisely one unmarked zero. Encircle this single unmarked zero and cross any other zero in its row to make an assignment to it. Continue until you’ve gone through all of the columns.

Step 4 – Perform the Optimal Test

• The present assignment is optimal if each row and column has exactly one encircled zero.
• The present assignment is not optimal if at least one row or column is missing an assignment (i.e., if at least one row or column is missing one encircled zero). Continue to step 5. Subtract the least cost element from all the entries in each column of the final cost matrix created in step 1 and ensure that each column has at least one zero.

Step 5 – Draw the least number of straight lines to cover all of the zeros as follows:

(a) Highlight the rows that aren’t assigned.

(b) Label the columns with zeros in marked rows (if they haven’t already been marked).

(c) Highlight the rows that have assignments in indicated columns (if they haven’t previously been marked).

(d) Continue with (b) and (c) until no further marking is needed.

(f) Simply draw the lines through all rows and columns that are not marked. If the number of these lines equals the order of the matrix, then the solution is optimal; otherwise, it is not.

Step 6 – Find the lowest cost factor that is not covered by the straight lines. Subtract this least-cost component from all the uncovered elements and add it to all the elements that are at the intersection of these straight lines, but leave the rest of the elements alone.

Step 7 – Continue with steps 1 – 6 until you’ve found the highest suitable assignment.

Hungarian Method Example

Use the Hungarian method to solve the given assignment problem stated in the table. The entries in the matrix represent each man’s processing time in hours.

\(\begin{array}{l}\begin{bmatrix} & I & II & III & IV & V \\1 & 20 & 15 & 18 & 20 & 25 \\2 & 18 & 20 & 12 & 14 & 15 \\3 & 21 & 23 & 25 & 27 & 25 \\4 & 17 & 18 & 21 & 23 & 20 \\5 & 18 & 18 & 16 & 19 & 20 \\\end{bmatrix}\end{array} \)

With 5 jobs and 5 men, the stated problem is balanced.

\(\begin{array}{l}A = \begin{bmatrix}20 & 15 & 18 & 20 & 25 \\18 & 20 & 12 & 14 & 15 \\21 & 23 & 25 & 27 & 25 \\17 & 18 & 21 & 23 & 20 \\18 & 18 & 16 & 19 & 20 \\\end{bmatrix}\end{array} \)

Subtract the lowest cost element in each row from all of the elements in the given cost matrix’s row. Make sure that each row has at least one zero.

\(\begin{array}{l}A = \begin{bmatrix}5 & 0 & 3 & 5 & 10 \\6 & 8 & 0 & 2 & 3 \\0 & 2 & 4 & 6 & 4 \\0 & 1 & 4 & 6 & 3 \\2 & 2 & 0 & 3 & 4 \\\end{bmatrix}\end{array} \)

Subtract the least cost element in each Column from all of the components in the given cost matrix’s Column. Check to see if each column has at least one zero.

\(\begin{array}{l}A = \begin{bmatrix}5 & 0 & 3 & 3 & 7 \\6 & 8 & 0 & 0 & 0 \\0 & 2 & 4 & 4 & 1 \\0 & 1 & 4 & 4 & 0 \\2 & 2 & 0 & 1 & 1 \\\end{bmatrix}\end{array} \)

When the zeros are assigned, we get the following:

The present assignment is optimal because each row and column contain precisely one encircled zero.

Where 1 to II, 2 to IV, 3 to I, 4 to V, and 5 to III are the best assignments.

Hence, z = 15 + 14 + 21 + 20 + 16 = 86 hours is the optimal time.

Practice Question on Hungarian Method

Use the Hungarian method to solve the following assignment problem shown in table. The matrix entries represent the time it takes for each job to be processed by each machine in hours.

\(\begin{array}{l}\begin{bmatrix}J/M & I & II & III & IV & V \\1 & 9 & 22 & 58 & 11 & 19 \\2 & 43 & 78 & 72 & 50 & 63 \\3 & 41 & 28 & 91 & 37 & 45 \\4 & 74 & 42 & 27 & 49 & 39 \\5 & 36 & 11 & 57 & 22 & 25 \\\end{bmatrix}\end{array} \)

Stay tuned to BYJU’S – The Learning App and download the app to explore all Maths-related topics.

Frequently Asked Questions on Hungarian Method

What is hungarian method.

The Hungarian method is defined as a combinatorial optimization technique that solves the assignment problems in polynomial time and foreshadowed subsequent primal–dual approaches.

What are the steps involved in Hungarian method?

The following is a quick overview of the Hungarian method: Step 1: Subtract the row minima. Step 2: Subtract the column minimums. Step 3: Use a limited number of lines to cover all zeros. Step 4: Add some more zeros to the equation.

What is the purpose of the Hungarian method?

When workers are assigned to certain activities based on cost, the Hungarian method is beneficial for identifying minimum costs.

Leave a Comment Cancel reply

Your Mobile number and Email id will not be published. Required fields are marked *

Request OTP on Voice Call

Post My Comment

Register with BYJU'S & Download Free PDFs

Register with byju's & watch live videos.

• Implementation of the Hungarian algorithm
• Connection to the Successive Shortest Path Algorithm
• Task examples
• Practice Problems

Hungarian algorithm for solving the assignment problem ¶

Statement of the assignment problem ¶.

There are several standard formulations of the assignment problem (all of which are essentially equivalent). Here are some of them:

There are $n$ jobs and $n$ workers. Each worker specifies the amount of money they expect for a particular job. Each worker can be assigned to only one job. The objective is to assign jobs to workers in a way that minimizes the total cost.

Given an $n \times n$ matrix $A$ , the task is to select one number from each row such that exactly one number is chosen from each column, and the sum of the selected numbers is minimized.

Given an $n \times n$ matrix $A$ , the task is to find a permutation $p$ of length $n$ such that the value $\sum A[i]\left[p[i]\right]$ is minimized.

Consider a complete bipartite graph with $n$ vertices per part, where each edge is assigned a weight. The objective is to find a perfect matching with the minimum total weight.

It is important to note that all the above scenarios are " square " problems, meaning both dimensions are always equal to $n$ . In practice, similar " rectangular " formulations are often encountered, where $n$ is not equal to $m$ , and the task is to select $\min(n,m)$ elements. However, it can be observed that a "rectangular" problem can always be transformed into a "square" problem by adding rows or columns with zero or infinite values, respectively.

We also note that by analogy with the search for a minimum solution, one can also pose the problem of finding a maximum solution. However, these two problems are equivalent to each other: it is enough to multiply all the weights by $-1$ .

Hungarian algorithm ¶

Historical reference ¶.

The algorithm was developed and published by Harold Kuhn in 1955. Kuhn himself gave it the name "Hungarian" because it was based on the earlier work by Hungarian mathematicians Dénes Kőnig and Jenő Egerváry. In 1957, James Munkres showed that this algorithm runs in (strictly) polynomial time, independently from the cost. Therefore, in literature, this algorithm is known not only as the "Hungarian", but also as the "Kuhn-Mankres algorithm" or "Mankres algorithm". However, it was recently discovered in 2006 that the same algorithm was invented a century before Kuhn by the German mathematician Carl Gustav Jacobi . His work, About the research of the order of a system of arbitrary ordinary differential equations , which was published posthumously in 1890, contained, among other findings, a polynomial algorithm for solving the assignment problem. Unfortunately, since the publication was in Latin, it went unnoticed among mathematicians.

It is also worth noting that Kuhn's original algorithm had an asymptotic complexity of $\mathcal{O}(n^4)$ , and only later Jack Edmonds and Richard Karp (and independently Tomizawa ) showed how to improve it to an asymptotic complexity of $\mathcal{O}(n^3)$ .

The $\mathcal{O}(n^4)$ algorithm ¶

To avoid ambiguity, we note right away that we are mainly concerned with the assignment problem in a matrix formulation (i.e., given a matrix $A$ , you need to select $n$ cells from it that are in different rows and columns). We index arrays starting with $1$ , i.e., for example, a matrix $A$ has indices $A[1 \dots n][1 \dots n]$ .

We will also assume that all numbers in matrix A are non-negative (if this is not the case, you can always make the matrix non-negative by adding some constant to all numbers).

Let's call a potential two arbitrary arrays of numbers $u[1 \ldots n]$ and $v[1 \ldots n]$ , such that the following condition is satisfied:

(As you can see, $u[i]$ corresponds to the $i$ -th row, and $v[j]$ corresponds to the $j$ -th column of the matrix).

Let's call the value $f$ of the potential the sum of its elements:

On one hand, it is easy to see that the cost of the desired solution $sol$ is not less than the value of any potential.

Lemma. $sol\geq f.$

The desired solution of the problem consists of $n$ cells of the matrix $A$ , so $u[i]+v[j]\leq A[i][j]$ for each of them. Since all the elements in $sol$ are in different rows and columns, summing these inequalities over all the selected $A[i][j]$ , you get $f$ on the left side of the inequality, and $sol$ on the right side.

On the other hand, it turns out that there is always a solution and a potential that turns this inequality into equality . The Hungarian algorithm described below will be a constructive proof of this fact. For now, let's just pay attention to the fact that if any solution has a cost equal to any potential, then this solution is optimal .

Let's fix some potential. Let's call an edge $(i,j)$ rigid if $u[i]+v[j]=A[i][j].$

Recall an alternative formulation of the assignment problem, using a bipartite graph. Denote with $H$ a bipartite graph composed only of rigid edges. The Hungarian algorithm will maintain, for the current potential, the maximum-number-of-edges matching $M$ of the graph $H$ . As soon as $M$ contains $n$ edges, then the solution to the problem will be just $M$ (after all, it will be a solution whose cost coincides with the value of a potential).

Let's proceed directly to the description of the algorithm .

Step 1. At the beginning, the potential is assumed to be zero ( $u[i]=v[i]=0$ for all $i$ ), and the matching $M$ is assumed to be empty.

Step 2. Further, at each step of the algorithm, we try, without changing the potential, to increase the cardinality of the current matching $M$ by one (recall that the matching is searched in the graph of rigid edges $H$ ). To do this, the usual Kuhn Algorithm for finding the maximum matching in bipartite graphs is used. Let us recall the algorithm here. All edges of the matching $M$ are oriented in the direction from the right part to the left one, and all other edges of the graph $H$ are oriented in the opposite direction.

Recall (from the terminology of searching for matchings) that a vertex is called saturated if an edge of the current matching is adjacent to it. A vertex that is not adjacent to any edge of the current matching is called unsaturated. A path of odd length, in which the first edge does not belong to the matching, and for all subsequent edges there is an alternating belonging to the matching (belongs/does not belong) - is called an augmenting path. From all unsaturated vertices in the left part, a depth-first or breadth-first traversal is started. If, as a result of the search, it was possible to reach an unsaturated vertex of the right part, we have found an augmenting path from the left part to the right one. If we include odd edges of the path and remove the even ones in the matching (i.e. include the first edge in the matching, exclude the second, include the third, etc.), then we will increase the matching cardinality by one.

If there was no augmenting path, then the current matching $M$ is maximal in the graph $H$ .

Step 3. If at the current step, it is not possible to increase the cardinality of the current matching, then a recalculation of the potential is performed in such a way that, at the next steps, there will be more opportunities to increase the matching.

Denote by $Z_1$ the set of vertices of the left part that were visited during the last traversal of Kuhn's algorithm, and through $Z_2$ the set of visited vertices of the right part.

Let's calculate the value $\Delta$ :

Lemma. $\Delta > 0.$

Suppose $\Delta=0$ . Then there exists a rigid edge $(i,j)$ with $i\in Z_1$ and $j\notin Z_2$ . It follows that the edge $(i,j)$ must be oriented from the right part to the left one, i.e. $(i,j)$ must be included in the matching $M$ . However, this is impossible, because we could not get to the saturated vertex $i$ except by going along the edge from j to i. So $\Delta > 0$ .

Now let's recalculate the potential in this way:

for all vertices $i\in Z_1$ , do $u[i] \gets u[i]+\Delta$ ,

for all vertices $j\in Z_2$ , do $v[j] \gets v[j]-\Delta$ .

Lemma. The resulting potential is still a correct potential.

We will show that, after recalculation, $u[i]+v[j]\leq A[i][j]$ for all $i,j$ . For all the elements of $A$ with $i\in Z_1$ and $j\in Z_2$ , the sum $u[i]+v[j]$ does not change, so the inequality remains true. For all the elements with $i\notin Z_1$ and $j\in Z_2$ , the sum $u[i]+v[j]$ decreases by $\Delta$ , so the inequality is still true. For the other elements whose $i\in Z_1$ and $j\notin Z_2$ , the sum increases, but the inequality is still preserved, since the value $\Delta$ is, by definition, the maximum increase that does not change the inequality.

Lemma. The old matching $M$ of rigid edges is valid, i.e. all edges of the matching will remain rigid.

For some rigid edge $(i,j)$ to stop being rigid as a result of a change in potential, it is necessary that equality $u[i] + v[j] = A[i][j]$ turns into inequality $u[i] + v[j] < A[i][j]$ . However, this can happen only when $i \notin Z_1$ and $j \in Z_2$ . But $i \notin Z_1$ implies that the edge $(i,j)$ could not be a matching edge.

Lemma. After each recalculation of the potential, the number of vertices reachable by the traversal, i.e. $|Z_1|+|Z_2|$ , strictly increases.

First, note that any vertex that was reachable before recalculation, is still reachable. Indeed, if some vertex is reachable, then there is some path from reachable vertices to it, starting from the unsaturated vertex of the left part; since for edges of the form $(i,j),\ i\in Z_1,\ j\in Z_2$ the sum $u[i]+v[j]$ does not change, this entire path will be preserved after changing the potential. Secondly, we show that after a recalculation, at least one new vertex will be reachable. This follows from the definition of $\Delta$ : the edge $(i,j)$ which $\Delta$ refers to will become rigid, so vertex $j$ will be reachable from vertex $i$ .

Due to the last lemma, no more than $n$ potential recalculations can occur before an augmenting path is found and the matching cardinality of $M$ is increased. Thus, sooner or later, a potential that corresponds to a perfect matching $M^*$ will be found, and $M^*$ will be the answer to the problem. If we talk about the complexity of the algorithm, then it is $\mathcal{O}(n^4)$ : in total there should be at most $n$ increases in matching, before each of which there are no more than $n$ potential recalculations, each of which is performed in time $\mathcal{O}(n^2)$ .

We will not give the implementation for the $\mathcal{O}(n^4)$ algorithm here, since it will turn out to be no shorter than the implementation for the $\mathcal{O}(n^3)$ one, described below.

The $\mathcal{O}(n^3)$ algorithm ¶

Now let's learn how to implement the same algorithm in $\mathcal{O}(n^3)$ (for rectangular problems $n \times m$ , $\mathcal{O}(n^2m)$ ).

The key idea is to consider matrix rows one by one , and not all at once. Thus, the algorithm described above will take the following form:

Consider the next row of the matrix $A$ .

While there is no increasing path starting in this row, recalculate the potential.

As soon as an augmenting path is found, propagate the matching along it (thus including the last edge in the matching), and restart from step 1 (to consider the next line).

To achieve the required complexity, it is necessary to implement steps 2-3, which are performed for each row of the matrix, in time $\mathcal{O}(n^2)$ (for rectangular problems in $\mathcal{O}(nm)$ ).

To do this, recall two facts proved above:

With a change in the potential, the vertices that were reachable by Kuhn's traversal will remain reachable.

In total, only $\mathcal{O}(n)$ recalculations of the potential could occur before an augmenting path was found.

From this follow these key ideas that allow us to achieve the required complexity:

To check for the presence of an augmenting path, there is no need to start the Kuhn traversal again after each potential recalculation. Instead, you can make the Kuhn traversal in an iterative form : after each recalculation of the potential, look at the added rigid edges and, if their left ends were reachable, mark their right ends reachable as well and continue the traversal from them.

Developing this idea further, we can present the algorithm as follows: at each step of the loop, the potential is recalculated. Subsequently, a column that has become reachable is identified (which will always exist as new reachable vertices emerge after every potential recalculation). If the column is unsaturated, an augmenting chain is discovered. Conversely, if the column is saturated, the matching row also becomes reachable.

To quickly recalculate the potential (faster than the $\mathcal{O}(n^2)$ naive version), you need to maintain auxiliary minima for each of the columns:

$minv[j]=\min_{i\in Z_1} A[i][j]-u[i]-v[j].$

It's easy to see that the desired value $\Delta$ is expressed in terms of them as follows:

$\Delta=\min_{j\notin Z_2} minv[j].$

Thus, finding $\Delta$ can now be done in $\mathcal{O}(n)$ .

It is necessary to update the array $minv$ when new visited rows appear. This can be done in $\mathcal{O}(n)$ for the added row (which adds up over all rows to $\mathcal{O}(n^2)$ ). It is also necessary to update the array $minv$ when recalculating the potential, which is also done in time $\mathcal{O}(n)$ ( $minv$ changes only for columns that have not yet been reached: namely, it decreases by $\Delta$ ).

Thus, the algorithm takes the following form: in the outer loop, we consider matrix rows one by one. Each row is processed in time $\mathcal{O}(n^2)$ , since only $\mathcal{O}(n)$ potential recalculations could occur (each in time $\mathcal{O}(n)$ ), and the array $minv$ is maintained in time $\mathcal{O}(n^2)$ ; Kuhn's algorithm will work in time $\mathcal{O}(n^2)$ (since it is presented in the form of $\mathcal{O}(n)$ iterations, each of which visits a new column).

The resulting complexity is $\mathcal{O}(n^3)$ or, if the problem is rectangular, $\mathcal{O}(n^2m)$ .

Implementation of the Hungarian algorithm ¶

The implementation below was developed by Andrey Lopatin several years ago. It is distinguished by amazing conciseness: the entire algorithm consists of 30 lines of code .

The implementation finds a solution for the rectangular matrix $A[1\dots n][1\dots m]$ , where $n\leq m$ . The matrix is ​1-based for convenience and code brevity: this implementation introduces a dummy zero row and zero column, which allows us to write many cycles in a general form, without additional checks.

Arrays $u[0 \ldots n]$ and $v[0 \ldots m]$ store potential. Initially, they are set to zero, which is consistent with a matrix of zero rows (Note that it is unimportant for this implementation whether or not the matrix $A$ contains negative numbers).

The array $p[0 \ldots m]$ contains a matching: for each column $j = 1 \ldots m$ , it stores the number $p[j]$ of the selected row (or $0$ if nothing has been selected yet). For the convenience of implementation, $p[0]$ is assumed to be equal to the number of the current row.

The array $minv[1 \ldots m]$ contains, for each column $j$ , the auxiliary minima necessary for a quick recalculation of the potential, as described above.

The array $way[1 \ldots m]$ contains information about where these minimums are reached so that we can later reconstruct the augmenting path. Note that, to reconstruct the path, it is sufficient to store only column values, since the row numbers can be taken from the matching (i.e., from the array $p$ ). Thus, $way[j]$ , for each column $j$ , contains the number of the previous column in the path (or $0$ if there is none).

The algorithm itself is an outer loop through the rows of the matrix , inside which the $i$ -th row of the matrix is ​​considered. The first do-while loop runs until a free column $j0$ is found. Each iteration of the loop marks visited a new column with the number $j0$ (calculated at the last iteration; and initially equal to zero - i.e. we start from a dummy column), as well as a new row $i0$ - adjacent to it in the matching (i.e. $p[j0]$ ; and initially when $j0=0$ the $i$ -th row is taken). Due to the appearance of a new visited row $i0$ , you need to recalculate the array $minv$ and $\Delta$ accordingly. If $\Delta$ is updated, then the column $j1$ becomes the minimum that has been reached (note that with such an implementation $\Delta$ could turn out to be equal to zero, which means that the potential cannot be changed at the current step: there is already a new reachable column). After that, the potential and the $minv$ array are recalculated. At the end of the "do-while" loop, we found an augmenting path ending in a column $j0$ that can be "unrolled" using the ancestor array $way$ .

The constant INF is "infinity", i.e. some number, obviously greater than all possible numbers in the input matrix $A$ .

To restore the answer in a more familiar form, i.e. finding for each row $i = 1 \ldots n$ the number $ans[i]$ of the column selected in it, can be done as follows:

The cost of the matching can simply be taken as the potential of the zero column (taken with the opposite sign). Indeed, as you can see from the code, $-v[0]$ contains the sum of all the values of $\Delta$ ​​, i.e. total change in potential. Although several values ​​​​of $u[i]$ and $v[j]$ could change at once, the total change in the potential is exactly equal to $\Delta$ , since until there is an augmenting path, the number of reachable rows is exactly one more than the number of the reachable columns (only the current row $i$ does not have a "pair" in the form of a visited column):

Connection to the Successive Shortest Path Algorithm ¶

The Hungarian algorithm can be seen as the Successive Shortest Path Algorithm , adapted for the assignment problem. Without going into the details, let's provide an intuition regarding the connection between them.

The Successive Path algorithm uses a modified version of Johnson's algorithm as reweighting technique. This one is divided into four steps:

• Use the Bellman-Ford algorithm, starting from the sink $s$ and, for each node, find the minimum weight $h(v)$ of a path from $s$ to $v$ .

For every step of the main algorithm:

• Reweight the edges of the original graph in this way: $w(u,v) \gets w(u,v)+h(u)-h(v)$ .
• Use Dijkstra 's algorithm to find the shortest-paths subgraph of the original network.
• Update potentials for the next iteration.

Given this description, we can observe that there is a strong analogy between $h(v)$ and potentials: it can be checked that they are equal up to a constant offset. In addition, it can be shown that, after reweighting, the set of all zero-weight edges represents the shortest-path subgraph where the main algorithm tries to increase the flow. This also happens in the Hungarian algorithm: we create a subgraph made of rigid edges (the ones for which the quantity $A[i][j]-u[i]-v[j]$ is zero), and we try to increase the size of the matching.

In step 4, all the $h(v)$ are updated: every time we modify the flow network, we should guarantee that the distances from the source are correct (otherwise, in the next iteration, Dijkstra's algorithm might fail). This sounds like the update performed on the potentials, but in this case, they are not equally incremented.

To deepen the understanding of potentials, refer to this article .

Task examples ¶

Here are a few examples related to the assignment problem, from very trivial to less obvious tasks:

Given a bipartite graph, it is required to find in it the maximum matching with the minimum weight (i.e., first of all, the size of the matching is maximized, and secondly, its cost is minimized). To solve it, we simply build an assignment problem, putting the number "infinity" in place of the missing edges. After that, we solve the problem with the Hungarian algorithm, and remove edges of infinite weight from the answer (they could enter the answer if the problem does not have a solution in the form of a perfect matching).

Given a bipartite graph, it is required to find in it the maximum matching with the maximum weight . The solution is again obvious, all weights must be multiplied by minus one.

The task of detecting moving objects in images : two images were taken, as a result of which two sets of coordinates were obtained. It is required to correlate the objects in the first and second images, i.e. determine for each point of the second image, which point of the first image it corresponded to. In this case, it is required to minimize the sum of distances between the compared points (i.e., we are looking for a solution in which the objects have taken the shortest path in total). To solve, we simply build and solve an assignment problem, where the weights of the edges are the Euclidean distances between points.

The task of detecting moving objects by locators : there are two locators that can't determine the position of an object in space, but only its direction. Both locators (located at different points) received information in the form of $n$ such directions. It is required to determine the position of objects, i.e. determine the expected positions of objects and their corresponding pairs of directions in such a way that the sum of distances from objects to direction rays is minimized. Solution: again, we simply build and solve the assignment problem, where the vertices of the left part are the $n$ directions from the first locator, the vertices of the right part are the $n$ directions from the second locator, and the weights of the edges are the distances between the corresponding rays.

Covering a directed acyclic graph with paths : given a directed acyclic graph, it is required to find the smallest number of paths (if equal, with the smallest total weight) so that each vertex of the graph lies in exactly one path. The solution is to build the corresponding bipartite graph from the given graph and find the maximum matching of the minimum weight in it. See separate article for more details.

Tree coloring book . Given a tree in which each vertex, except for leaves, has exactly $k-1$ children. It is required to choose for each vertex one of the $k$ colors available so that no two adjacent vertices have the same color. In addition, for each vertex and each color, the cost of painting this vertex with this color is known, and it is required to minimize the total cost. To solve this problem, we use dynamic programming. Namely, let's learn how to calculate the value $d[v][c]$ , where $v$ is the vertex number, $c$ is the color number, and the value $d[v][c]$ itself is the minimum cost needed to color all the vertices in the subtree rooted at $v$ , and the vertex $v$ itself with color $c$ . To calculate such a value $d[v][c]$ , it is necessary to distribute the remaining $k-1$ colors among the children of the vertex $v$ , and for this, it is necessary to build and solve the assignment problem (in which the vertices of the left part are colors, the vertices of the right part are children, and the weights of the edges are the corresponding values of $d$ ). Thus, each value $d[v][c]$ is calculated using the solution of the assignment problem, which ultimately gives the asymptotic $\mathcal{O}(nk^4)$ .

If, in the assignment problem, the weights are not on the edges, but on the vertices, and only on the vertices of the same part , then it's not necessary to use the Hungarian algorithm: just sort the vertices by weight and run the usual Kuhn algorithm (for more details, see a separate article ).

Consider the following special case . Let each vertex of the left part be assigned some number $\alpha[i]$ , and each vertex of the right part $\beta[j]$ . Let the weight of any edge $(i,j)$ be equal to $\alpha[i]\cdot \beta[j]$ (the numbers $\alpha[i]$ and $\beta[j]$ are known). Solve the assignment problem. To solve it without the Hungarian algorithm, we first consider the case when both parts have two vertices. In this case, as you can easily see, it is better to connect the vertices in the reverse order: connect the vertex with the smaller $\alpha[i]$ to the vertex with the larger $\beta[j]$ . This rule can be easily generalized to an arbitrary number of vertices: you need to sort the vertices of the first part in increasing order of $\alpha[i]$ values, the second part in decreasing order of $\beta[j]$ values, and connect the vertices in pairs in that order. Thus, we obtain a solution with complexity of $\mathcal{O}(n\log n)$ .

The Problem of Potentials . Given a matrix $A[1 \ldots n][1 \ldots m]$ , it is required to find two arrays $u[1 \ldots n]$ and $v[1 \ldots m]$ such that, for any $i$ and $j$ , $u[i] + v[j] \leq a[i][j]$ and the sum of elements of arrays $u$ and $v$ is maximum. Knowing the Hungarian algorithm, the solution to this problem will not be difficult: the Hungarian algorithm just finds such a potential $u, v$ that satisfies the condition of the problem. On the other hand, without knowledge of the Hungarian algorithm, it seems almost impossible to solve such a problem.

This task is also called the dual problem of the assignment problem: minimizing the total cost of the assignment is equivalent to maximizing the sum of the potentials.

Literature ¶

Ravindra Ahuja, Thomas Magnanti, James Orlin. Network Flows [1993]

Harold Kuhn. The Hungarian Method for the Assignment Problem [1955]

James Munkres. Algorithms for Assignment and Transportation Problems [1957]

Practice Problems ¶

UVA - Crime Wave - The Sequel

UVA - Warehouse

SGU - Beloved Sons

UVA - The Great Wall Game

UVA - Jogging Trails

• Alessandro Minisini (92.97%)
• Oleksandr Kulkov (7.03%)

• DSA Tutorial
• Data Structures
• Linked List
• Dynamic Programming
• Binary Tree
• Binary Search Tree
• Divide & Conquer
• Mathematical
• Backtracking
• Branch and Bound
• Pattern Searching

Hungarian Algorithm for Assignment Problem | Set 2 (Implementation)

Given a 2D array , arr of size N*N where arr[i][j] denotes the cost to complete the j th job by the i th worker. Any worker can be assigned to perform any job. The task is to assign the jobs such that exactly one worker can perform exactly one job in such a way that the total cost of the assignment is minimized.

Input: arr[][] = {{3, 5}, {10, 1}} Output: 4 Explanation: The optimal assignment is to assign job 1 to the 1st worker, job 2 to the 2nd worker. Hence, the optimal cost is 3 + 1 = 4. Input: arr[][] = {{2500, 4000, 3500}, {4000, 6000, 3500}, {2000, 4000, 2500}} Output: 4 Explanation: The optimal assignment is to assign job 2 to the 1st worker, job 3 to the 2nd worker and job 1 to the 3rd worker. Hence, the optimal cost is 4000 + 3500 + 2000 = 9500.

Different approaches to solve this problem are discussed in this article .

Approach: The idea is to use the Hungarian Algorithm to solve this problem. The algorithm is as follows:

• For each row of the matrix, find the smallest element and subtract it from every element in its row.
• Repeat the step 1 for all columns.
• Cover all zeros in the matrix using the minimum number of horizontal and vertical lines.
• Test for Optimality : If the minimum number of covering lines is N , an optimal assignment is possible. Else if lines are lesser than N , an optimal assignment is not found and must proceed to step 5.
• Determine the smallest entry not covered by any line. Subtract this entry from each uncovered row, and then add it to each covered column. Return to step 3.

Consider an example to understand the approach:

Let the 2D array be: 2500 4000 3500 4000 6000 3500 2000 4000 2500 Step 1: Subtract minimum of every row. 2500, 3500 and 2000 are subtracted from rows 1, 2 and 3 respectively. 0   1500  1000 500  2500   0 0   2000  500 Step 2: Subtract minimum of every column. 0, 1500 and 0 are subtracted from columns 1, 2 and 3 respectively. 0    0   1000 500  1000   0 0   500  500 Step 3: Cover all zeroes with minimum number of horizontal and vertical lines. Step 4: Since we need 3 lines to cover all zeroes, the optimal assignment is found.   2500   4000  3500  4000  6000   3500   2000  4000  2500 So the optimal cost is 4000 + 3500 + 2000 = 9500

For implementing the above algorithm, the idea is to use the max_cost_assignment() function defined in the dlib library . This function is an implementation of the Hungarian algorithm (also known as the Kuhn-Munkres algorithm) which runs in O(N 3 ) time. It solves the optimal assignment problem. 

Below is the implementation of the above approach:

Time Complexity: O(N 3 ) Auxiliary Space: O(N 2 )

Please Login to comment...

Similar reads.

• How to Get a Free SSL Certificate
• Best SSL Certificates Provider in India
• Elon Musk's xAI releases Grok-2 AI assistant
• What is OpenAI SearchGPT? How it works and How to Get it?
• Content Improvement League 2024: From Good To A Great Article

Improve your Coding Skills with Practice

What kind of Experience do you want to share?

Index     Assignment problem     Hungarian algorithm     Solve online    

The Hungarian algorithm: An example

We consider an example where four jobs (J1, J2, J3, and J4) need to be executed by four workers (W1, W2, W3, and W4), one job per worker. The matrix below shows the cost of assigning a certain worker to a certain job. The objective is to minimize the total cost of the assignment.

Below we will explain the Hungarian algorithm using this example. Note that a general description of the algorithm can be found here .

Step 1: Subtract row minima

We start with subtracting the row minimum from each row. The smallest element in the first row is, for example, 69. Therefore, we substract 69 from each element in the first row. The resulting matrix is:

Step 2: Subtract column minima

Similarly, we subtract the column minimum from each column, giving the following matrix:

Step 3: Cover all zeros with a minimum number of lines

We will now determine the minimum number of lines (horizontal or vertical) that are required to cover all zeros in the matrix. All zeros can be covered using 3 lines:

Step 4: Create additional zeros

First, we find that the smallest uncovered number is 6. We subtract this number from all uncovered elements and add it to all elements that are covered twice. This results in the following matrix:

Now we return to Step 3.

Again, We determine the minimum number of lines required to cover all zeros in the matrix. Now there are 4 lines required:

Because the number of lines required (4) equals the size of the matrix ( n =4), an optimal assignment exists among the zeros in the matrix. Therefore, the algorithm stops.

The optimal assignment

The following zeros cover an optimal assignment:

This corresponds to the following optimal assignment in the original cost matrix:

Thus, worker 1 should perform job 3, worker 2 job 2, worker 3 job 1, and worker 4 should perform job 4. The total cost of this optimal assignment is to 69 + 37 + 11 + 23 = 140.

Solve your own problem online

HungarianAlgorithm.com © 2013-2024

Quantitative Techniques: Theory and Problems by P. C. Tulsian, Vishal Pandey

Get full access to Quantitative Techniques: Theory and Problems and 60K+ other titles, with a free 10-day trial of O'Reilly.

There are also live events, courses curated by job role, and more.

HUNGARIAN METHOD

Although an assignment problem can be formulated as a linear programming problem, it is solved by a special method known as Hungarian Method because of its special structure. If the time of completion or the costs corresponding to every assignment is written down in a matrix form, it is referred to as a Cost matrix. The Hungarian Method is based on the principle that if a constant is added to every element of a row and/or a column of cost matrix, the optimum solution of the resulting assignment problem is the same as the original problem and vice versa. The original cost matrix can be reduced to another cost matrix by adding constants to the elements of rows and columns where the total cost or the total completion time of an ...

Get Quantitative Techniques: Theory and Problems now with the O’Reilly learning platform.

O’Reilly members experience books, live events, courses curated by job role, and more from O’Reilly and nearly 200 top publishers.

Don’t leave empty-handed

Get Mark Richards’s Software Architecture Patterns ebook to better understand how to design components—and how they should interact.

It’s yours, free.

Check it out now on O’Reilly

Dive in for free with a 10-day trial of the O’Reilly learning platform—then explore all the other resources our members count on to build skills and solve problems every day.

Hungarian Method: Assignment Problem

Hungarian Method is an efficient method for solving assignment problems .

This method is based on the following principle:

• If a constant is added to, or subtracted from, every element of a row and/or a column of the given cost matrix of an assignment problem, the resulting assignment problem has the same optimal solution as the original problem.

Hungarian Algorithm

The objective of this section is to examine a computational method - an algorithm - for deriving solutions to the assignment problems. The following steps summarize the approach:

Steps in Hungarian Method

1. Identify the minimum element in each row and subtract it from every element of that row.

2. Identify the minimum element in each column and subtract it from every element of that column.

3. Make the assignments for the reduced matrix obtained from steps 1 and 2 in the following way:

• For every zero that becomes assigned, cross out (X) all other zeros in the same row and the same column.
• If for a row and a column, there are two or more zeros and one cannot be chosen by inspection, then you are at liberty to choose the cell arbitrarily for assignment.

4. An optimal assignment is found, if the number of assigned cells equals the number of rows (and columns). In case you have chosen a zero cell arbitrarily, there may be alternate optimal solutions. If no optimal solution is found, go to step 5.

5. Draw the minimum number of vertical and horizontal lines necessary to cover all the zeros in the reduced matrix obtained from step 3 by adopting the following procedure:

• Mark all the rows that do not have assignments.
• Mark all the columns (not already marked) which have zeros in the marked rows.
• Mark all the rows (not already marked) that have assignments in marked columns.
• Repeat steps 5 (i) to (iii) until no more rows or columns can be marked.
• Draw straight lines through all unmarked rows and marked columns.

You can also draw the minimum number of lines by inspection.

6. Select the smallest element from all the uncovered elements. Subtract this smallest element from all the uncovered elements and add it to the elements, which lie at the intersection of two lines. Thus, we obtain another reduced matrix for fresh assignment.

7. Go to step 3 and repeat the procedure until you arrive at an optimal assignment.

For the time being we assume that number of jobs is equal to number of machines or persons. Later in the chapter, we will remove this restrictive assumption and consider a special case where no. of facilities and tasks are not equal.

Share This Article

Operations Research Simplified Back Next

Goal programming Linear programming Simplex Method Transportation Problem

Stack Exchange Network

Stack Exchange network consists of 183 Q&A communities including Stack Overflow , the largest, most trusted online community for developers to learn, share their knowledge, and build their careers.

Q&A for work

Connect and share knowledge within a single location that is structured and easy to search.

Difference between solving Assignment Problem using the Hungarian Method vs. LP

When trying to solve for assignments given a cost matrix, what is the difference between

using Scipy's linear_sum_assignment function (which I think uses the Hungarian method)

describing the LP problem using a objective function with many boolean variables, add in the appropriate constraints and send it to a solver, such as through scipy.optimize.linprog ?

Is the later method slower than Hungarian method's O(N^3) but allows for more constraints to be added?

• linear-programming
• combinatorial-optimization
• assignment-problem

• $\begingroup$ The main difference between a mathematical model and a heuristic algorithm to solve a specific problem is more likely to prove optimality rather feasibility. Now, one can decide which one to be selected in order to satisfy needs. $\endgroup$ –  A.Omidi Commented Mar 20, 2022 at 7:54
• 4 $\begingroup$ @A.Omidi the Hungarian method is an exact algorithm $\endgroup$ –  fontanf Commented Mar 20, 2022 at 8:26
• $\begingroup$ @fontanf, you are right. What I said was to compare the exact and heuristic methods and it is not specific to Hungarian alg. Thanks for your hint. $\endgroup$ –  A.Omidi Commented Mar 20, 2022 at 10:32

The main differences probably are that there is a somewhat large overhead you have to pay when solving the AP as a linear program: You have to build an LP model and ship it to a solver. In addition, an LP solver is a generalist. It solves all LP problems and focus in development is to be fast on average on all LPs and also to be fast-ish in the pathological cases.

When using the Hungarian method, you do not build a model, you just pass the cost matrix to a tailored algorithm. You will then use an algorithm developed for that specific problem to solve it. Hence, it will most likely solve it faster since it is a specialist.

So if you want to solve an AP you should probably use the tailored algorithm. If you plan on extending your model to handle other more general constraints as well, you might need the LP after all.

Edit: From a simple test in Python, my assumption is confirmed in this specific setup (which is to the advantage of the Hungarian method, I believe). The set up is as follows:

• A size is chosen in $n\in \{5,10,\dots,500\}$
• A cost matrix is generated. Each coefficient $c_{ij}$ is generated as a uniformly distributed integer in the range $[250,999]$ .
• The instance is solved using both linear_sum_assignment and as a linear program. The solution time is measured as wall clock time and only the time spent by linear_sum_assignment and the solve function is timed (not building the LP and not not generating the instance)

For each size, I have generated and solved ten instances, and I report the average time only.

And then there is of course the "but". I am not a ninja in Python and I have used pyomo for modelling the LPs. I believe that pyomo is known to be slow-ish whenbuilding models, hence I have only timed the solver.solve(model) part of the code - not building the model. There is however possibly a hugh overhead cost coming from pyomo translating the model to "gurobian" (I use gurobi as solver).

• 1 $\begingroup$ Do you have some benchmark results to support this claim? Intuitively, I would have thought that the Hungarian method would be much slower in practice $\endgroup$ –  fontanf Commented Mar 20, 2022 at 7:39
• $\begingroup$ @fontanf I only have anecdotal proof from past experiments. Maybe an LP solver can work faster for repeated solves, where you exploit that the model is already build and basis info is available. But I honestly don't know. $\endgroup$ –  Sune Commented Mar 20, 2022 at 9:37
• 1 $\begingroup$ It might be the case that the Hungarian method is faster for small problems (due to the overhead Sune mentioned for setting up an LP model) while simplex (or dual simplex, or maybe barrier) might be faster for large models because the setup cost is "amortized" better. (I'm just speculating here.) $\endgroup$ –  prubin ♦ Commented Mar 20, 2022 at 15:30
• 2 $\begingroup$ The Hungarian algorithm is, of course, O(n^3) for fully dense assignment problems. I don't know if there is a simplex bound explicitly for assignments. Simplex is exponential in the worst case and linear in variables plus constraints (n^2 + 2n here) in practice. But assignments are highly degenerate (n positive basics out of 2n rows). Dual simplex may fare better than primal. Hungarian is all integer for integer costs, whereas a standard simplex code won't be unless it knows to detect that in preprocessing. That may lead to some overhead for linear algebra. Ha, an idea for a class project! $\endgroup$ –  mjsaltzman Commented Mar 20, 2022 at 17:04
• 2 $\begingroup$ Just for the sake of completeness, here 's the same with gurobipy instead of Pyomo. On my machine, all LPs (n = 500) are solved in less than a second compared to roughly 4 seconds with Pyomo. $\endgroup$ –  joni Commented Mar 21, 2022 at 16:21

Your Answer

Sign up or log in, post as a guest.

Required, but never shown

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy .

Not the answer you're looking for? Browse other questions tagged linear-programming combinatorial-optimization assignment-problem or ask your own question .

• Featured on Meta
• Bringing clarity to status tag usage on meta sites
• We've made changes to our Terms of Service & Privacy Policy - July 2024
• Announcing a change to the data-dump process

Hot Network Questions

• How long does it take to achieve buoyancy in a body of water?
• Using conditionals within \tl_put_right from latex3 explsyntax
• Could someone tell me what this part of an A320 is called in English?
• Trying to find an old book (fantasy or scifi?) in which the protagonist and their romantic partner live in opposite directions in time
• How to run tests for all flows in an org?
• Does Vexing Bauble counter taxed 0 mana spells?
• Unsure about reduced state of entangled bits after measurement in different basis
• Integral concerning the floor function
• What prevents a browser from saving and tracking passwords entered to a site?
• If inflation/cost of living is such a complex difficult problem, then why has the price of drugs been absoultly perfectly stable my whole life?
• Using "no" at the end of a statement instead of "isn't it"?
• How to reply to reviewers who ask for more work by responding that the paper is complete as it stands?
• Parody of Fables About Authenticity
• Two way ANOVA or two way repeat measurement ANOVA
• How can I get an Edge's Bevel Weight attribute value via Python?
• Has the US said why electing judges is bad in Mexico but good in the US?
• Is it possible to have a planet that's gaslike in some areas and rocky in others?
• My school wants me to download an SSL certificate to connect to WiFi. Can I just avoid doing anything private while on the WiFi?
• How can these humans cross the ocean(s) at the first possible chance?
• Reusing own code at work without losing licence
• Is having negative voltages on a MOSFET gate a good idea?
• Distinctive form of "לאהוב ל-" instead of "לאהוב את"
• Is this a new result about hexagon?
• My Hydraulic brakes are seizing up and I have tried everything. Help

The assignment problem as a cooperative game

• Published: 28 August 2024

Cite this article

• V. V. Morozov 1 &
• S. I. Romanov 1  

We consider the problem of optimal distribution of assignments between workers on machines. Workers can exchange machines with the aim of saving the total time to execute the assignments, compared to their initial distribution. It is proved that the corresponding cooperative game has a non-empty c -core and is totally balanced. Based on the Hungarian method of solving the assignment problem, an algorithm for constructing an imputation from the c -core is formulated. In the case of an integer characteristic function, the algorithm constructs an imputation from the c -core with integer components. This result is generalized to the case of arbitrary totally balanced games of three and four persons, as well as for symmetric games.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Subscribe and save.

• Get 10 units per month
• Download Article/Chapter or eBook
• 1 Unit = 1 Article or 1 Chapter
• Cancel anytime

Price includes VAT (Russian Federation)

Instant access to the full article PDF.

Rent this article via DeepDyve

Institutional subscriptions

Mazalov, V.V.: Matematicheskaya teoriya igr i prilozheniya [in Russian] (Mathematical game theory and applications). Lan, Saint-Petersburg (2010)

Google Scholar  

Peleg, B., Sudhölter, P.: Introduction to the theory of cooperative games. Springer-Verlag, Berlin (2007)

Bondareva, O.N.: Nekotorye primeneniya metodov lineynogo programmirovaniya k teorii kooperativnyh igr [in Russian] (Some applications of linear programming methods to the theory of cooperative games). Probl. Cybern. 10 , 119–140 (1963)

Shapley, L.S.: On balanced sets and cores. RM-4601-PR. The Rand Corporation, Santa Monica, California (1965)

Charnes, A., Kortanek, K.: On balanced sets, cores and programming. Cah. Centre Etudes Rech. Oper. 9 (1), 32–43 (1967)

MathSciNet   Google Scholar  

Vasin, A.A., Morozov, V.V.: Teoriya igr i modeli matematicheskoy ekonomiki [in Russian] (Game theory and models of mathematical economics). MAX Press, Moscow (2005)

Peleg, B.: An inductive method for constructing minimal balanced collections of finite sets. Nav. Res. Logist. Quart. 12 (2), 155–162 (1965)

Article   MathSciNet   Google Scholar  

Raiser, G.D.: Combinatorial mathematics. Mir, Moscow (1966)

Kuhn, H.W.: The Hungarian method for the assignment problem. Nav. Res. Logist. Quart. 2 (1-2), 155–162 (1955)

Download references

Author information

Authors and affiliations.

Faculty of Computational Mathematics and Cybernetics, Lomonosov Moscow State University, Moscow, Russian Federation

V. V. Morozov & S. I. Romanov

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to V. V. Morozov .

Additional information

Translated from Prikladnaya Matematika i Informatika , No. 74, pp. 19–26, 2023

This article is a translation of the original article published in Russian in the journal Prikladnaya Matematika i Informatika . The translation was done with the help of an artificial intelligence machine translation tool, and subsequently reviewed and revised by an expert with knowledge of the field. Springer Nature works continuously to further the development of tools for the production of journals, books and on the related technologies to support the authors.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Morozov, V.V., Romanov, S.I. The assignment problem as a cooperative game. Comput Math Model (2024). https://doi.org/10.1007/s10598-024-09600-0

Download citation

Received : 06 May 2023

Accepted : 12 October 2023

Published : 28 August 2024

DOI : https://doi.org/10.1007/s10598-024-09600-0

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Optimal assignment problem
• Core of a cooperative game
• Hungarian method
• Directed multigraph
• Find a journal
• Publish with us
• Track your research