This course covers the essential information that every serious programmer needs to know about algorithms and data structures, with emphasis on applications and scientific performance analysis of Java implementations. Part I covers elementary data structures, sorting, and searching algorithms. Part II focuses on graph- and string-processing algorithms.
Maxflow Applications
Loading...
From the course by Princeton University
Algorithms, Part II
442 ratings
From the lesson
Maximum Flow and Minimum Cut
In this lecture we introduce the maximum flow and minimum cut problems. We begin with the Ford−Fulkerson algorithm. To analyze its correctness, we establish the maxflow−mincut theorem. Next, we consider an efficient implementation of the Ford−Fulkerson algorithm, using the shortest augmenting path rule. Finally, we consider applications, including bipartite matching and baseball elimination.
Meet the Instructors
Robert SedgewickWilliam O. Baker *39 Professor of Computer Science
Computer Science
Kevin WayneSenior Lecturer
Computer Science
Let's finish up by looking at some applications of maxflow like shortest
paths maxflow is a very widely-applicable problem solving model. And it is really
important to recognize at this stage, We've looked at a lot of algorithms for
solving specific problems. And they're important problems.
And it's important to have efficient algorithms for solving them.
But when you have something like maxflow or shortest paths the, the importance that
we attach to them is really magnified by the idea that they have this property
that, that they're a very general way to state a problem and we have many, many
practical situations that we can cast in terms of these problems.
We looked at arbitrage coming or reducing down to a shortest paths problem.
And we'll look at a bunch of problems that don't seem to be related at all, that can
be modeled as maxflow problems. So, they're extremely important because
they're problem solving models that work for a broad variety of important
applications. Number one that wouldn't be any good if we
didn't have efficient algorithms for solving them.
But we do have efficient algorithms for solving them.
And so, that magnifies their importance. And that's why people work so hard on
finding efficient algorithms for solving these problems.
And we'll talk about that as well in a minute.
maxflow We saw an image processing algorithm called syncarving for shortest
path. There's another one called segmentation
for maxflow. Again, if you have an image and you have one vertex per pixel you have
a huge, huge graph. And you have many explicit huge graphs and
we've talked about those types of things. But there are other things where the graph
is, is an abstraction that it gets involved in a model of the abstract graph
and the maxflow problem. Its maybe a bit surprising at first, and
we'll look at a couple examples of that to illustrate the point.
Over here is a medical example having to do with it.
That's the, the image processing one on a medical example to help identify some
important part of a medical image. So, we'll look at a, at a couple examples
to that the idea of a general problem solving model that, once we have an
efficient algorithm, then we can think about using the problem solving model.
And later on, we'll see that this, this concept of a general problem solving model
has really profound implications and we'll be looking at that later on.
So, let's just look at a, at a couple of examples.
Here's one called the bipartite matching problem.
So you have this is a bit of an idealized situation.
But it works in more messy, real life situations, too.
So, there's n jobs out there and n students apply for them.
And we'll use a small example where there's five students and five jobs.
But, of course, in the real world, this can be huge.
Now during hiring season, the students go out and apply for the jobs and they each
get a bunch of offers. So looking at it from a student's point of
view. Alice gets offers from Adobe, Amazon, and
Google. Adobe makes offers to Alice, Bob, and
Carol, and like that. So, this is an association between
students and jobs. Everybody gets several offers.
And in question is well, it would be good if everybody got a job, right?
And the question is, is there some way for everybody to get a job.
That's called the bipartite matching problem.
And it comes up in lots of forms directly related to graph processing.
Now, we could study and people do study algorithms for explicitly solving this
particular problem. But what I want to emphasize is that
actually, maxflow is a reasonable model for it.
So, we can use our efficient maxflow implementation to get it solved.
We don't have to come up with a specialized algorithm for this problem.
So, in terms of graphs, it's called the bipartite matching problem,
Given a bipartite graph, find a perfect matching.
And a bipartite graph is one where you have two sets of vertices, in this case,
one to corresponding to students and another corresponding to companies.
And you have every edge in the graph goes from one type of vertex to other, the
other type of vertex. And a matching in the graph is a set of
edges that are disjoint that disconnects two vertices but that's it.
And so, in this case, there's a perfect matching works out that if Alice takes the
Google job and Bob takes the Adobe job and Carol takes the Facebook job and like
that, then everybody gets a job. So, that's a perfect matching.
But you can also have a situation where that's not possible.
So let's look at how to formulate, How to, well, the one thing is, how do we
find the matching? And then the other thing is, is there one?
So this is easy to formulate as a matching network flow problem.
That's what this diagram shows. So, what we'll do is we'll create our
source and target vertices. We'll have one vertex for each student.
One vertex for each company in the flow network.
And we'll add a capacity one edge from s to each student.
And a capacity one edge from t to from each company to t and then it doesn't
matter what the capacity. We'll add an infinite capacity edge from
each student to each job offer. And then, we'll ask for a maximum flow in
this graph. So, you can see that the flow, every
augmenting path has to go from s to a student to a company to t and so, the flow
will give us the match and let's see how it works.
This is a, a one to one correspondence between perfect matchings and bipartite
graphs, and integer value maxflows. so, in this case, there's a flow of value five.
And that flow gives us the matching immediately.
So what the mean cut tells us if, if there's a no perfect matching, explain
why. So, here's an example that maybe could
have happened with the job offers. And when the we're algorithm terminates it
terminates with a cut we're the, a cut of the bipartite graph, which separates two,
four, and five from seven and ten. And essentially the cut tells us that
students in two, four, or five can only be matched to companies seven and ten.
You could see all the edges from two, four, five go to seven and ten, so you
have two companies and three students. So, there's no way that everybody can be
matched, somebody's gonna be left out. So that's a the students, so that they'll
be a mean cut, and the s will be the students on the s side and t will be the
companies on the s side and if it's bigger than, s is bigger than t, then I can't
have a matching. So in this case at, there's only, four
jobs and somebody is going to be left out. It's also interesting to trace through
what happens with the maxflow algorithm on bipartite graphs like this.
Essentially augmenting paths or usually forward edges makes some matching.
And then if it's possible to find a path that undoes some matching.
It, zig zags through, undoing matching and trying other ones to find a way through to
the target. But if there's no perfect matching,
there'll be a mean cut. And that one will explain why.
So, that's a problem, The bipartite matching problem that we can
model as a maxflow algorithm and just use our existing code to solve it.
Here's another one that's even further away.
It doesn't seem to have a graph at all, but it does.
It's called the baseball elimination problem.
In this is again, just to show the breadth of applicability of the maxflow model.
It's interesting at certain times of year, you get near the of the baseball season
and often you'll hear in the news, or see in the paper, or see in the web, that your
team is mathematically eliminated. Actually most of the time, they don't
really get that right because they don't do the computation that we're going to
talk about next. Sometimes, it's easy this is an example
where it's easy. So we've got four teams, they already have
this win-loss record and this is the number of games to play.
So in this case Montreal has only three games to play. so the best they could do
is win 80. Ag, but Atlanta has already got 83 wins so
there's no way Montreal is going to win. So, that's a mathematical elimination that
anyone could figure out. Usually the newspaper will get that one
right. So but sometimes it's more complicated if
you look, say, this case. So Philly has 80 wins, three games to
play. So the best they can do is 83 wins.
So that's interesting. But the thing is that Atlanta has a bunch
of games against, it's got six games against the Mets.
And either Atlanta wins one of them, which would give Atlanta 84 wins, or the Mets
win all of them, in which case, they get 84 wins.
Either way, Philadelphia is mathematically eliminated.
That's a bit more complicated decision about which team wins.
The thing is and there's many more complicated situations that show up. And
the observation, just from these two easy examples, is that you can't figure out
who's mathematically eliminated without knowing the full schedule of games.
It depends, not only on how many games were already won, how many are left to
play, but it depends on the schedule and who's playing who.
And usually, your average sportswriter is not going to do that computation without a
computer. And I hope that one of you becomes a
sportswriter and puts this in for the future for us.
So let's look at a more complicated situation.
So this is the American League East awhile ago near the end of the season.
And question is which teams are mathematically eliminated and which ones
aren't. Now in this case it turn's out that the,
this is pretty far from the end of the season actually.
These 27 games to finish. And this is a proof here that Detroit is
mathematically eliminated. But it's a pretty complicated argument and
well, you can, you can reason it out with arithmetic.
The tough part is to figure out this set of teams here are.
So what we're going to see is that you can do a maxflow computation to figure out
this sets of teams. And this, let's just look at it for this
example. So, at this point, Detroit is
mathematically eliminated. And so it's got 27 games to play, so it
could in theory, win 76 of the games. Now but the logic that will convince you
that they are eliminated is that if you take the four teams the other four teams
and add up all their wins there's 278 of them.
And you look at the remaining games there's 27.
So somebody's gotta win every one of those games.
The total number of games won for that set of The teams is 305, and if you divide by
four. It means the average is 76.25.
And right there is a proof that one of them is got to win 77 games.
That takes a little thought, but if you have the four teams, then from the
remaining games, you can figure out that Detroit is mathematically eliminated.
But the key is, how do we find those four teams.
And the answer, as I've already said, is it's maxflow. and so this is a maxflow
network that can be used to solve the baseball elimination problem.
So the intuition is that, that you have a source vertex and you have what happens in
the remaining games flowing from the source to the target.
So here we're trying to prove that team four is we're trying to decide if team
four is eliminated or not. That's Detroit, in this case.
And so, what we need is a vertex for each pair of vertices that are not the team
we're interested in. And so, that's going to relate to all the
remaining games because these are the pairs of teams.
And then you have an edge from the source to each one of those vertices.
And the capacity of the edge is the number of games left between those two.
So that's on one end. And then, you have a vertex for each team.
And then what we do for each one of these pair of vertices,
We put infinite capacity edges to the two teams involved.
So, the flow is going to be an integer flow, so some of it will go one way and
some of it will go the, the other way. But then, for each of the teams what we're
going to do is, make sure that they don't win more games than team four, the team
we're interested in. So, we'll put this upper bound on the flow
here that we won't let the numbers of wins get better than what our team of interest,
team four can do. And the fact is that if you compute a
maxflow of this you can convince yourself, that if you can fill this network up
going, going from s in, in the maxflow Then team four, team four is not going to
be eliminated. Nobody's going to get more wins than team
four. And so the way to solve the baseball
elimination problem is to run maxflow on this network, and the mean cut will give
the set of keys, it's our, mean cut will give the set of teams that you needed in
this calculation to figure out to prove to a friend that, or to a sportswriter that
the team you're interested in is, is eliminated.
An interesting application of maxflow. Again, we just take our problem, use it to
build a network solve the problem on the network using our existing code and
translate that solution to a solution to our original problem.
That's called reduction and it's a very important technique that we're going to
use we're going to talk about it in some detail later on.
So now we come to the theory of maxflow algorithms.
This is ,, an even hotter area than minimum spanning tree and shortest paths
that we've looked at before and that it's a very frustrating situation for
theoretical computer scientists. And that we have this relatively
straightforward to state algorithm and we have this all, this design freedom,
forward focus in algorithm. There's lots and lots of ways that we
could try to find augmenting paths and there's even other methods that don't use
forward focus and that are almost as simple. and the question is, how difficult
is it to solve the maxflow problem? And there's literally hundreds of papers
in the scientific literature that are oriented at trying to solve this problem.
Now, again the, the theoretical computer scientists are trying to find an algorithm
that's guaranteed to work well in the worst case.
So, they're just counting the number of edges that the algorithms examine in the
worst case. But when related to practical graphs these
are very, very conservative upper bounds. And the real performance is going to be
totally different. So, you can't use these to compare the
performance of a given algorithm. The performance of a given algorithm
really depends on the characteristic of networks.
But still, there's a huge gap between the best algorithms that we know.
In a most recent one was discovered just this,
This year that can guarantee e squared over log e, number of edges examined to
small compared to say, shortest augmenting path which is e cubed essentially.
And that's, that's fine, but actually in practice, the running time of many of the
algorithms seems to be relatively small factor of e,
And no one can prove that there might not exist an algorithm, or no one has proved
yet, that there might not exist an algorithm that gets the job done in linear
time. So one of the exciting things about
studying the field of algorithms is there's still room to find, to discover
interesting and innovative algorithms that could have a huge practical impact.
Because we have algorithms that won't run well on practical networks.
Lots and lots of important practical applications use them.
And if someone, someone or discovered, to discover a fast, practical, guaranteed
linear time algorithm, it would immediately have huge impact.
So that's the first warning was worst case order of growth.
You're not going to compare algorithms in practice.
And there's plenty of research papers out there that have done empirical studies on
the maxflow algorithms for realistic networks in the so-called ath, best so far
in practice is known as the push-relabel method with gap relabeling, which runs in
time e square v, where e is the number of edges.
And, again, even that in practical networks is going to run faster than that.
So, there's numerous research challenges still to be addressed in studying the
maxflow problem. There's plenty of practitioners that are
using the codes like the one's that we've shown and, and variations to try to solve
a huge real maxflow mincut problems and trying to get them done in linear time.
There's many theoretical computer scientist that are trying to prove that
there exists or not exists, a maxflow algorithm that is guaranteed to run in
linear time, no matter what the input. There's many, many people doing and
there's still a great deal to be learned. It's a fine example of why it's exciting
to be working in the field of algorithms. There's, an opportunity for new knowledge
still available and many people are still working on them.
Explore our Catalog
Join for free and get personalized recommendations, updates and offers.
Coursera provides universal access to the world’s best education,
partnering with top universities and organizations to offer courses online.
© 2018 Coursera Inc. All rights reserved.
