LAPJV - Jonker-Volgenant Algorithm for Linear Assignment Problem V3.0

A Matlab implementation of the Jonker-Volgenant algorithm solving LAPs.

Description

The Jonker-Volgenant algorithm is much faster than the famous Hungarian algorithm for the Linear Assignment Problem (LAP). This Matlab implementation is modified from the original C++ code made by Roy Jonker, one of the inventors of the algorithm. It is about 10 times faster than the munkres code (v2.2) of the author. It can solve a 1000 x 1000 problem in about 3 seconds in a normal Intel Centrino processor.

V1.1 returns the dual variables and the reduced cost matrix as well.
V1.2 can deal with nonsquare assignment problems.
V2.0 is faster for problems with a large range of costs.
V2.1 includes an option to change the cost resolution to improve performance for some problems.
v2.2 removes a small bug to avoid NAN for 1x1 case.
v2.3 removes a small bug to handle a cost matrix with all inf's.
v2.4 fixes a bug associated with resolution to address the known problem of the algorithm.
v3.0 fixes a bug introduced since v2.0.
Reference:
R. Jonker and A. Volgenant, "A shortest augmenting path algorithm for dense and spare linear assignment problems", Computing, Vol. 38, pp. 325-340, 1987.

Acknowledgements

Munkres Assignment Algorithm and Hungarian Algorithm For Linear Assignment Problems (V2.3) inspired this file.

This file inspired K Best Assignment Algorithm and Faster Jonker Volgenant Assignment Algorithm.

• MATLAB release: MATLAB 7.10 (R2010a)

[Don Natale]

Hi Yi,

I'm running the example you have in the comments of your code:

% Example 2: 1000 x 1000 random data
%{
n=1000;
A=randn(n)./rand(n);
tic
[a,b]=lapjv(A);
toc % about 0.5 seconds
%}

And it's taking substantially longer than a few seconds to solve, more like 40 seconds. This is on a fairly new Intel core i5-2400 3.10GHz processor with 64-bit Windows 7 and 16 Gb RAM.

In your great amount of experience optimizing your code here, could you suggest a reason why this may be taking so long to compute?

[Fabio]

Hi,
This code looks very interesting.
I am using v3.0 and I possibly hit a bug/problem. It seems to me that, for different cost matrices (coming from the underlying problem with a varying size) the solution switches between being in row mode and col mode.
I don't see any parameter/switch to control this behavior, is it intended to be like that? how can I get the solution always in term of e.g. col indices? Fabio

[Bill Otto]

v3.0 seems to give correct result every time, with a 4x speed advantage over Munkres for a 100 by 100 matrix. Thanks for the great routine, Yi.

[Bill Otto]

Yi, It seems to fail for 2x2 (0.0022% of the time) and 3x3 cost matrices about 0.014% of random cases. For 4x4, it hasn't failed in over 1 million cases, but that doesn't mean that it won't in some larger sample. That means it's a real tough bug to track down, I would guess. Bill

[Bill Otto]

Yi, I have tried Jen's example below, and it appears to be a bona fide bug. Bill

[Jacob]

I have the following error, can anybody help me here?

??? Error: File: lapjv.m Line: 146 Column: 7
Expression or statement is incorrect--possibly unbalanced (, {, or [.

[Jen]

It seems like the code returns the wrong solution for this cost matrix:

costs =
[0.2593 0.0697 0.7715;
0.1671 0.3699 0.0671;
0.1117 0.1462 0.0611]

[a, b] = lapjv(costs);
a = [2 1 3]
b = 0.2979

The correct solution would be a = [2 3 1] with b = 0.2486.

[Kristofer Kusano]

Thanks for the post, this worked great.

[joshua vogelstein]

this is a really great piece of code, thanks!
i noticed that "Faster Jonker-Volgenant Assignment Algorithm" seems to have a nice feature which is not yet implemented in this version? do you have plans to include it? many thanks, joshua

[Ming]

Great

[Yi Cao]

Thank you Dmitri, the bug has been fixed now.

Yi

[Dmitri Kamenetsky]

I think I found a bug in lapjv when Inf values are used. In the following lapjv gives the wrong answer (compared to munkres):

A=magic(5);
A(1,3)=Inf
[munkresAssignment munkresCost]=munkres(A)
[lapjvAssignment lapjvCost]=lapjv(A)

The problem goes away when Inf is replaced with a large value:

A(1,3)=1000
[munkresAssignment munkresCost]=munkres(A)
[lapjvAssignment lapjvCost]=lapjv(A)

[Andy]

My post was deleted somehow

I am having a problem with bad assignments. The data I am using is two sets of 8x2 centroids specified by (row,col). The first set is measured data and tsecond data is predicted via alpha-beta filter. I using the assignemtn algorithm to match up masured data with existing filter pipelines, but even for a simple scenario such as:

meas =
[122.9178,122.9178;122.9178,202.9178;0,0;0,0;0,0;0,0;0,0;0,0];
pred =
[122.4879,121.0182;0,0;0,0;0,0;0,0;0,0;0,0;0,0];

i get
rowsol = [2;3;6;7;8;1;4;5],

wich is obviously not right. Even with Munkres i get
rowsol_munkres = [2;1;3;4;5;6;7;8],

which is still off.

Any suggestions?

[Andy]

Sorry here is the costmat code:

costMat = zeros(n,n);

%Populate nxn error matrix
%--------------------------------
for j = 1:n
for k = 1:n
costMat(j,k) = ( x_pred(j)-x_meas(k) )^2 + ( y_pred(j)-y_meas(k) )^2;
end
end

[Eric Trautmann]

I've been using both of your Munkres and JV LAP algorithm implementations. JV is much faster for a single matrix, as expected, but I also found that when I used the JV algorithm within the Murty k-best assignment algorithm, it was much slower than your implementation of Munkres. When I dug into this, I found that the JV implementation does not cull out rows or columns all marked as inf. When I made this modification (copied the validRow/validCols lines from munkres.m), it sped up the algorithm by a factor of ~7.

[Andrew McFarland]

Alexander,
Thanks for the help...I tried this out and it did help a lot, but it is still taking much longer with my dataset (for example, 80X80 cost matrix takes ~150 seconds to solve) vs the example ones. Any ideas--maybe normalizing of some sort?

[Alexander Kosenkov]

Andrew McFarland, you should specify second parameter - minimum resoulution for your data. If several values are very close to each other, then it hangs because of a minor bug in programming.

I made a patch to the function - originally EPS function was used incorrectly. I hope, Mr. Cao will fix this in next releases.

Index: lapjv.m
@@ -68,7 +68,7 @@
if nargin<2
- resolution=eps;
+ resolution=4 * eps(max(max(A)));
end
@@ -164,7 +164,7 @@
i0 = colsol(j1);
- if usubmin - umin > eps
+ if usubmin - umin > resolution
% change the reduction of the minmum column to increase the

[Andrew McFarland]

Great code--I have found some interesting behavior. When I run the "standard" examples like A=rand(1000,1000) for the cost matrix, solves in seconds. However, I have an assignment problem (for PCB assignment) and when I run a 500X500 cost matrix it takes more than a day to solve. I have tried scaling the matrix values and also randomly shuffling the matrix elements and still doesn't reduce the solver time. Is there something inherent in the structure of the cost matrix that can cause the solution to take longer??

[KS]

 

[Yi Cao]

KS,

The code is updated with more clear explaination. Wish this helps.

Yi

[KS]

What does N stand for in the function parameters? Thanks.

[Yi Cao]

Roberto,

Thanks for pointing this out. Initially, a full Inf cost matrix is not considered. Now, this has been corrected in v2.3.

Yi

[Roberto Olmi]

I've noticed a little bug for scalar input Inf ( lapjv(inf) ). The error occurs at line 85.

[Roberto Olmi]

Very usefull code!! Compared to the version 1.1 the version 2.2 seems to be also capable to manage Inf values in the cost matrix. Yi, is it correct?

[Jiangmin zhang]

i think matlab should take this code into itself as a build-in function

[Amir]

anyone has lapjv v1.2?

[Immanuel Weber]

Hi Yi,
that was fast! Okay, that's the solution I want to avoid, as the task in my case is very time critical and resizing an existing 2d array appears to be very unattractive. But nevertheless thanks for your information and your good code.

Immanuel

[Yi Cao]

Immanuel,

Thanks for pointing out the bug. It has been removed in v2.2. In terms of rectangular matriices, I did not know these two references you mentioned. I just used a simple approach to fill the cost matrix with zeros to square it.

Yi

[Immanuel Weber]

Hi Yi,
as I used your Hungarian-Code before and I want to implement Jonkers&Volgenant's algorithm for rectangular matrices in C, your code was a nice finding. Do you use any of the algorithm modifications presented in Volgenant's papers "Linear and semi-assignment problems, a core oriented approach"[1996] and/or "Solving the Rectangular assignment problem and applications"[2010] to solve rectangular matrices? Or do you use a different, maybe simplier to understand approach?
I'm trying to implement the modifications presented in the last one, but I'm having difficulties with some of the lines in the pseudo code.
In addition I want to point out a little bug, which rarely will occur, but if you use 1x1 matrix as an input for your code, it will compute a NaN cost.

Immanuel

[Yi Cao]

Hi, Kishore,

No, this is not an error. This is a new feature of Matlab since 2009(a?), which indicates that the corresponding output arguments are not needed. If you use an older version, you need to replace ~ with a dummy variable.

Yi

[Kishore Mosaliganti]

Hi,

There is an error in line 85 in the downloaded code. It is not running.
[~,c]=min(v1);

Should the ~ be r ?

[Amir]

it is a wonderful code. I have some questions which directly or indirectly are related to your code. I would really appreciate if you could find some time and answer my questions:

1. If I want to exclude some assignments such as i-->i assignment, I consider Cii as Big-M. Does the value of Big-M affect the computing time? What is the best value?
2. If I have a matrix whose elements are random integers from the following interval [0 10000] and another one whose elements are from [0 100], does the computing time differ? My experience is that the first one is lengthier.
3. Do you have any experience working with a sparse matrix?
Everybody claims the computing time has to decrease. However, I have a different experience. It remains more or less the same. I think the main reason can be referred to the above item. I think solving a matrix with larger element with LAPJV is lengthier.
4. Can I have negative elements?
5. Is LAPJV v1.1 faster than LAPJV v1.2? Because, in that one you don't calculate dual variables and reduced cost matrix.

[Chaoran Du]

Hi Yi, I've got the m-file, thanks a lot!!!

Chaoran

[Yi Cao]

Chaoran,

The page has not been updated. You can check the 'updates' list for another 22 Jul 2010 item to ensure the file has been updated.

Yi

[Chaoran Du]

Hi Yi,

I still can't download the updated version and I don't know if there is something wrong here. Do you mind email me the updated m-file? Thank you very much for all your help!

Email: [e-mail]

[Yi Cao]

Chaoran,

I reloaded the file now. Wish this time it works.

Yi

[Chaoran Du]

Yi, I download the file but the m-file is the same as V1.1, and it shows that the m-file was modified on 19th July while license.txt was modified today. Could you please check the file?

Really appreciate all your help!

Chaoran

[Yi Cao]

Chaoran,

Yes, it has been posted and should be online within a day.

Yi

[Chaoran Du]

Fantastic!
But I can't find V1.2, have you posted it yet?
Thanks again for all your help!

Chaoran

[Yi Cao]

Chaoran,

The new version (V1.2) has the ability to deal with nonsquare cases.

Yi

[Chaoran Du]

Thank you for your code. Its really useful.
Do you know how to solve an asymmetric assignment problem by using the Jonker-Volgenant algorithm? (assign n people to m jobs and n<m)

[Yi Cao]

Amir,

In the new version (V1.1), dual variables are returned.

Yi

[Amir]

it is a great code. how can i get the dual variables?

[Amir Hossein]

 

[Turkay YILDIZ]

Thank you. Great algorithm and matlab code. I would also very much like to see this algorithm's solution for TSP and other problems, if it's possible.