Santa Routing and Heuristics in Operations Research

As a child, I always wondered how Santa Claus could visit all children of the world overnight. I now know that he must have a very optimized route plan. And this tells me one thing: Santa Claus is a math genius!

Santa's Vehicle Routing Problem

For simplicity, I only put some of the cities Santa needs to visit… And by no way do I mean that English kids have been too bad for Santa’s visit! Also, I can’t resist giving you this link to CGPGrey’s awesome video on Santa, even though it has nothing to do with vehicle routing.
What does math have to do with Santa?

In the last decades, similar problems called vehicle routing problems have become crucial for many companies, like Fedex and UPS. And these are so difficult, that companies asked help from applied mathematicians! Since then, applied mathematicians have enjoyed the beautifully combinatorial structure of these problems, as well as the millions of dollars companies have been willing to pay for answers! Let’s see what these mathematicians have come up with! This will draw a global picture of the world of operations research, which, more generally, aims at solving complex combinatorial industrial problems through mathematics.

I’m very thankful for Marilène Cherkesly for explaining me all the different technics in this article and helping me design the outline of this article!

Santa’s Vehicle Routing Problem

If there were no constraint on Santa Claus’ journey, the visiting-all-children problem would be a Traveling Salesman Problem (TSP): It would consist in searching for the shortest path visiting all cities and coming back to the original point. This class of problem has become iconic in the theory of computation, as it was proved to be very hard to solved!

So what’s a vehicle routing problem?

To make Santa Claus’ problem a vehicle routing problem, we merely have to assume that Santa Claus’ reindeers cannot carry too heavy quantities. So, for instance, they can only carry presents for 5 cities. This will force Santa to come back to his toy factory several times to reload his sleigh. Typically, a Santa’s journey on the night of December 24th could look as follows:

Example of Santa's Route

To have a chance to make it on time, it’s crucial for Santa to minimize the total distance of his journey. This corresponds to searching for a shortest overall journey. And that’s precisely what the Vehicle Routing Problem (VRP) is about.

Note that the assumption that reindeers can visit 5 cities only greatly simplifies the problem, as we now know that Santa will need 4 and only 4 tours. In the more common capacitated vehicle routing problem (CVRP), reindeers can carry a certain weight and different cities need to be delivered different amounts of toys, depending for instance on their sizes. In CVRP, the priority is to minimize the number of tours, which is even harder than our example where tours have a strong structure.
So how can it be solved?

The VRP is actually known to as hard as the TSP. This means that, when the number of cities isn’t small, there is virtually no hope to solve the VRP exactly in a reasonable amount of time, even with tomorrow’s greatest computer power. In fact, so far, no best solution is known for problems with more than a few hundreds of cities to visit!

What I’m saying here assumes that $P\neq NP$, which is the dominant belief in the complexity theorist community. In fact, I’m even assuming $BQP \neq NP$, where $BQP$ corresponds to the class of problems solvable efficiently with quantum computers. Find out more about $BQP$ in my article on probabilistic algorithms.

So, rather than solving the VRP exactly, operational researchers have been designing clever technics to find nearly optimal solutions in a reasonable amount of time. These technics are called heuristics.

Greedy Algorithms

Greedy Santa

The most straightforward heuristics is the greedy algorithm.

What’s a greedy algorithm?

It consists in choosing the next best step without much thinking, like a greedy Santa who is given a bunch of candies to choose from. He’d repeatedly eat his favorite one among the remaining candies, until he’s full… or until Mother Christmas tells him to stop!

It sounds like a great way to choose the best candies!

It is! In many cases, like for Santa’s candies, the greedy algorithm turns out to be optimal!

But how would that work for VRP?

A simple greedy algorithm would be to get Santa to the closest unvisited city.

Can you illustrate?

Sure! Starting in Santa’s workshop in Rovaniemi, and using this link, you can determine that the closest city is Moscow, at nearly 1,300 km. Then, the closest city to Moscow is Tehran. Then, the next closest cities to visit are Cairo, Paris and Marrakesh. Now that Santa has visited 5 cities, he must get back to Rovaniemi to pick up presents for other cities. Then, he goes on with his visits. Eventually, here’s the Santa’s journey we get:

Greedy Algorithm

Such a whole journey is called a solution. This is not to be confused with the optimal solution which would be the best Santa’s journey. Here, the solution found by the greedy algorithm is 124,800 kilometers long. Can we do better… And how?

Local Search

One way to do better is to improve what we have.

Isn’t it the only way?

Hummm… Yes. But what I mean by that is that we can try to make small modifications to the best solution we’ve found, and see if these modifications lead to an improvement. This technic is called local search. It’s pretty much what people use when they play the colder-warmer game, where each step is warmer when it improves one’s location.

Getting Warmer

So what kind of modifications are we talking about here?

It’s up to you to invent them! Experiments characterized a simple and efficient kind of modification called 2-opt. It consists in removing two arcs and reconnecting cities of these arcs in a consistent way.

Can you give an example?

Sure! Let’s remove the arcs New York – Mexico and Vancouver – Lima. Then, we have two possible reconstructions. The New York – Lima and Vancouver – Mexico reconstruction breaks Santa’s tour so it’s not acceptable. Conversely, the New York – Vancouver and Mexico – Lima reconstruction is consistent with Santa’s tour:

2-Opt

Note that the paths we removed add up to 11,600 km, while these we’ve added in the consistent reconstruction add up to 8,200km. Thus, by modifying Santa’s 2nd tour, we’ve reduced the total distance by 3,400km!

You chose to remove 2 arcs that crossed… Was that on purpose?

Great observation! By removing 2 arcs that crossed in the 2-opt modification, I was sure to greatly improve the current solution! Let me apply that to all crossing arcs. This is what we get:

Greedy + 2-Opt

Note that 2-opt on all crossing arcs may not work as we consider constraints such as pick-up and delivery or time windows.

Network of Solutions

Cool! But why is this method called this a local search?

To fully understand what is meant by local search, let me introduce the network of solutions. This network connects two solutions if a 2-opt modification enables to move from one to the other. Typically, the following figure displays a few of the solutions of the larger network of solutions:

Network of Santa's Journeys

Any solution is connected to a bunch of others, which form what its neighborhood. And since we only looked at 2-opt modifications, our local search corresponds to exploring the neighborhood of best solutions found. So, the search is dubbed local, because we’re looking for better solutions only in the neighborhood of the current best solution.

In fact, let’s pause a bit on this network of solutions. Let’s forget about the details of the network and, rather, let’s ponder upon the global structure of the network of solutions…

What do you mean?

Let’s look at the network as a cloud of solutions with connections between some of these solutions. When we apply a local search to this cloud, links can only be taken in certain direction: From the worse to the better solution. Thus, links actually have an orientation. In graph theoretical terms, this means that we actually have a directed (acyclic) graph of Santa’s journeys. And a local search will simply consist in taking a path in this graph until we face a dead-end.

Large Network of Solutions

This network of solution is an image I always try to keep in mind while doing optimization. It’s applicable to a huge class of optimization technics, including the state-of-the-art and fundamental linear programming, and to a lesser extent, the gradient methods. Indeed, in gradient methods, the graph is sort of dynamics as it depends on the step we take which usually decrease in time, and a good understanding of how we move from one solution to another is better reached with topology.
But I guess it’s not sure that we’ll get to the actual optimal solution with a local search…

I’d even add that it’s highly unlikely we will. One thing I can guarantee though is that we will reach a local optimum.

What’s a local optimum?

A local optimum is a solution which is better than any of its neighborhood. In other words, no arrow points out of a local optimum. So, crucially, it’s impossible with local search to get out of a local optimum. An that may be quite a problem, since a local optimum can be very bad compared to the global optimum.

Let me guess… The global optimum is a solution better than all others, right?

Exactly! And there’s worse than the fact that our algorithm may not lead us to the global optimum: Depending on the initial solution, it may simply be impossible for local searches to get to the global optimum, as there may be no path from the initial solution to the global optimum! This remark has led to new ideas for optimization.

Adaptive Large Neighborhood Search

Building upon this remark, applied mathematicians have then invented the Adaptive Large Neighborhood Search (ALNS). The idea of ALNS is to make radical modifications of the current solution to move away from local optima. Sort of like you sometimes need a revolution to really change the world…

What kind of radical modifications can we do in VRP?

Typically, we’d only keep half of the arrows of the current solution, and try to fill the gaps using a basic method. It’s common to include randomness both in construction and destruction to make sure that we will rebuild a very different solution.

Can you illustrate?

Sure. Let’s be clever and keep the arrows which seem relevant. Then let’s assign randomly colors to these arrows, and reconstruct afterwards how we can. Here’s what I got:

Adaptive Large Neighborhood Search

Note that, by now, we’ve improved the greedy solution (which was already not too bad) by over 10%! As turnovers of companies who require such models add up to billions of dollars, just think about how much we may save them… and how much they would owe us! It’s no surprise that some companies specialized in optimization have millions of turnover!

For good performance of ALNS, the level of destruction needs to be managed accordingly to the success of previous local search attempts. Note that there are several variants, like Variable Neighborhood Search (VNS) and Iterated Local Search (ILS).

Simulated Annealing and Tabu

But I guess that it’s quite unlikely for a radical modifications to improve the current solution, right?

Indeed. This led to simulated annealing. The idea is to still carry out radical modifications even though they are worse than the current best solution, and as long as there is hope they will lead to good solutions after local searches. However, after more and more iterations, we may become more and more demanding and eventually require a strict improvement from the current best solution.

Metallurgy

Why is it called simulated annealing?

By analogy with metal annealing! In a hot metal, particles still jiggle around at all kinds of energy levels. Importantly, as long as the metal is hot, particles can leave low energy state and get to a higher energy one. But as the metal cools down, particles reach lower energy states and have more trouble leaving them. Eventually, they rest at a low energy state which is often lower than the one they’d reach had the metal been frozen instantly.

So, eventually, in simulated annealing, no worse solution would be accepted, right?

Exactly! Here’s how to visualize simulated annealing used combined to local search and ALNS:

Simulated Annealing and ALNS Links

Given the arrows of simulated annealing, a path may walk back on its footsteps. To avoid that, the Tabu search has been introduced. It simply consists in remembering the last footsteps made to makes sure we never come back to a solution we already explored.

Genetic Algorithms

Father Christmas + Sinterklaas

In our example, combining all the heuristics discussed so far would yield nearly unbeatable solutions. But, as VRP models are made more realistic, they become much harder to solve, even heuristically! Examples of realistic constraints we may want to add include differences in quantity of toys required for cities, time windows to visit cities or reindeer required rests. For constraints like these, even slight modifications may greatly modify the quality of a solution. This leads to an explosion of the number of local optima. For this reason, the methods introduced so far can turn out quite limited.

So what can be done then?

One great idea came from evolutionary biology: Genetic Algorithms. The core idea lies in the fact that a mixture of two parents can make a kid much better than the average of the parents. An example of that is how crossing English Father Christmas with Dutch Sinterklaas has given birth to the world-renowned Santa Claus.

How would we do that in VRP?

The key to powerful genetic algorithms lies in the way we cross the parents’ genes. Researchers have come up with plenty of ways to do so, as listed in this article. Let me consider one to illustrate.

How does that work?

First, we represent solutions as the ordered sequence of all visited cities. Then, the father solution yields half of the cities to the kid, all placed exactly as they appear in the father’s ordering of cities. The gaps are then filled by the other half of the cities, in the order in which they appear in the mother solution.

Hummm… Can you give an example?

Sure. Let’s mix the greedy solution and the ALNS solution we have found earlier:

Genetic Algorithm

It actually looks unlikely that the kid is any good…

He needs education! Funnily, this term has been used to applying local search to the kid to make it an interesting solution!

Even with education, it doesn’t seem likely that the kid becomes the global optimum…

But, at least, he’ll contribute to a greater diversity of solutions. Indeed, the problem with local search is that the intensification leads to solutions which all look the same. It’d be a bit like having all physicists working on string theory alone. Sure, string theory is promising and it probably deserves intensification… But if all physicists looked in a single direction, they’d be missing out on a lot of other less promising but eventually more fruitful leads! Conversely, many breakthroughs in History have resulted from a clever mix of good ideas, like how Newton merged Galileo’s laws of motions on Earth with Kepler’s laws of planetary motions!

So a balance needs to be met between diversification and intensification, right?

Exactly! A classical approach consists in maintaining a pool of diverse not-too-bad solutions, called population. This population can be initially generated by a greedy randomized adaptive search procedure (GRASP). Then, two solutions of the population are picked and crossed over. Their kid is inserted in the population and replaces an unfit or non-original solution of the population. The choice of the solution to remove is called selection, and it regulates the level of diversification. This is represented below:

One Iteration in Genetic Algorithms

This sounds and looks nice, but do genetic algorithms really work?

Amazingly, since a paper by Christian Prins in 2004, genetic approaches have surpassed local searches in many problems.

Why did this success only occurred recently? Is it only now that researchers have thought of that?

Actually, genetic algorithms have been introduced as soon as the 1950s! But, crucially, computing power has greatly increased since, and you actually need to have this much greater computing capacity to get genetic algorithms running at their full potential!

So, do you think genetic algorithms will be ruling the future of optimization?

I’m sure that their ideas will be useful. But I rather believe in even more advanced optimization technics.

Integer Programming and Cuts

Like most combinatorial optimization problem, the VRP can be written as an integer linear program.

What’s an integer linear program?

It’s an optimization problems whose variables are all integers and which have a linear structure. More precisely, if you forget the fact that variables must take integer values, then you should be left with a linear program. This linear program has a linear objective function and linear constraints.

I’ve written a pretty complete series of articles on linear and integer programming (linear programming, duality, simplex methods and integer programming).
Hummm… What would an integer program for VRP look like?

Let’s encode whether Santa goes from city $i$ to city $j$ in tour number $k$ by binary variables $x_{ijk}$. More, precisely $x_{ijk}$ equals 1 if Santa does, and 0 otherwise.

Xijk

Now, to have an integer program, we also need an objective function and constraints.

What’s the objective function?

It’s what we want to minimize (or maximize). So, in our case, it’s the total distance of Santa’s journey. Crucially, it can be deduced from the variables. Indeed, denoting $d_{ij}$ between cities $i$ and $j$, the term $d_{ij} x_{ijk}$ equals $d_{ij}$ if Santa goes from $i$ to $j$ in tour $k$, and 0 otherwise. Thus, by summing over all tours $k$ and over all pairs of cities $i$ and $j$, we add up all the distances of intercity trips of Santa: That’s also the total distance of Santa’s journey! So, VRP consists in finding the variables $x_{ijk}$ which minimize the sum of terms $d_{ij}x_{ijk}$ over all values of $i$, $j$ and $k$.

Wait… Now it looks like always having $x_{ijk} = 0$ is the optimal solution!

You’re right! And that’s because our integer program so far includes solution which are inconsistent with the actual VRP. We now need to add constraints to make the integer program consistent with the problem it’s supposed to solve.

We need to add constraints? Like what?

Santa must go through all cities. In particular, if $i$ is a city which is not Santa’s, he must go through it exactly once. This implies that he must also leave it once. So, there must be one and only one variable $x_{ijk}$ which equals 1. Thus, all variables $x_{ijk}$ must add up to 1. So, for all cities $i$ different from Santa’s, we’ll add the constraint that the sum of $x_{ijk}$ equals 1.

But that’s still not enough, right?

Indeed. Let’s add the following constraints:

  • Santa goes and leaves its home at most once in each tour (0 if he makes no visit in the tour).
  • In each tour, through each city, there must be as many incoming arcs as outgoing ones.
  • A tour cannot be made of more than 5 cities other than Santa’s (and thus 6 arcs).

Now, given these constraints, it’s still possible to have inconsistent loops like below:

A Solution with a Subtour

Such a solution is said to have a subtour. To avoid these, a common method is to add subtour elimination constraints. They are based on the two following remarks. First, for any subset $S$ of cities not including Santa’s, there must be at least one arc going out of the subset. Thus, this outgoing arc comes from a city which has no outgoing arc which remains in the subset $S$. Second, the number of arcs within the subset $S$ equals the number of cities of the subset with an outgoing arc remaining in the subset $S$. Thus, by combining the two remarks, we deduce that the number of arcs within $S$ cannot equal or exceed the number of cities of the subset!

Is that it? Are there still more constraints to add?

Nope! We’re done! Here’s the complete integer programming formulation, with the mathematical translation of our constraints added on the right:

Integer Programming

Note that index 0 of cities is Santa’s toy factory city Rovaniemi, and the subset $S$ must not include 0.

This formulation is nearly perfect to put it directly into an integer program optimizer and quickly get a great solution!

What do you mean by “nearly”?

There’s one trouble though… The integer program above has an exponential number of constraints! Indeed, there are as many subtour elimination constraints as the number of subsets of the set of cities. And that’s bad, very bad…

Is it that bad?

If there are 100 cities, then there are $2^{100} \approx 10^{30}$ subtour elimination constraints! That’s more than the size of the web, which is only about $10^{18}$ bits! So, you can’t even write down the subtour elimination constraints for a VRP of 100 cities on the whole web!

This kind of make the formulation above useless, doesn’t it?

Fortunately, applied mathematicians have more than a few tricks up their sleeves! They have proposed to start the optimization without subtour elimination constraints. Then, for each solution found, we add the subtour elimination constraints the solution does not verify. If there is no such subtour elimination constraint, then this means that our solution is consistent… And is the global optimum too! This technic is called cutting planes, and has turned out to be today’s state-of-the-art technique for many problems. In fact, other constraints than subtour eliminations can be used in cutting plane methods to efficiently reach quality solutions.

Let’s Conclude

So, would you now agree with Santa’s being a math genius?

I guess… Unless his elves are the geniuses who plan the routing for him!

Haha… Now, I should say that there’s one last more state-of-the-art optimization technic I have not mentioned here. And yet, it’s the specialty of my research lab! It’s called column generation, and exploits the structure of the integer program to write it differently, using the Dantzig-Wolfe decomposition. Find out more with this article!

Do these more advanced technics mean that local searches and genetic algorithms are pointless?

Not at all. First, because other problems don’t have a structure as rich as the VRP does, which means that integer programming and column generation may not work for them, or would involve too much work and tests to get right. Second, and more importantly, these other methods can actually be used in combination with more advanced technics. This is particularly the case with column generation which decomposes the problem in smaller subproblems which may require local search or genetic algorithms! Overall, a clever balance of all these technics seems to be key for powerful optimization.

Cool! Also, you mentioned something about making these models more realistic…

Yes! By now, there have been a huge number of variants of the VRP. Among others are VRPPD (where cities correspond to pick-up or deliveries, and capacities must be modeled accordingly), LIFO-VRPPD (where the order in which you load and unload goods in trucks matter), VRPTW (where each city must be visited within a certain time window), SDVRP (where deliveries for a customer can be split in several trucks), HVRP (where the fleet of vehicles is heterogeneous), SVRP and RVRP (with uncertainties where we minimize the average case or the worst case). And, these can be combined, yielding problems like LIFO-SDVRPPDTW… As you can imagine, I like to tease my VRP colleagues about how they create new problems by just adding letters to the acronym!

No idea of gift? Send this article to your nerdy friends! Merry Christmas and Happy New Year! With a special dedication to all my vehicle routing researcher friends: Marilène, Charles, Claudio, Heiko, Remy, Sana, Vincent, Guillaume, François, Guy and those I forget!

More on Science4All

Optimization by Linear Programming Optimization by Linear Programming
By Lê Nguyên Hoang | Updated:2016-02 | Views: 7928
Operations Research deals with optimizing industrial systems. Those systems can be very complex and their modeling may require the use of hundreds, thousands or even millions of variables. Optimizing over millions of variables may seem impossible, but it can be done if the optimization problem has a linear structure. Learn more on this linear structure and optimization solutions!

Optimization by Integer Programming Optimization by Integer Programming
By Lê Nguyên Hoang | Updated:2016-01 | Prerequisites: Optimization by Linear Programming | Views: 8925
Integer programming is arguably the greatest achievement of applied mathematics. Half of the time, it's what's used to solve real-world problems!

Marriage Assignment Problem and Variants Marriage Assignment Problem and Variants
By Lê Nguyên Hoang | Updated:2016-02 | Views: 7323
Marriage problems consist in matching boys and girls. They are a very interesting class of problems that include assignment problems, which have a very large range of applications. Additional results have been found for variants which include the introduction of boys and girls' preferences, and cases where people cannot be categorized into two groups.

P versus NP: A Crucial Open Problem P versus NP: A Crucial Open Problem
By Lê Nguyên Hoang | Updated:2016-01 | Views: 7406
P=NP is probably today's most crucial open problem. Not only is it a very theoretical question in computer science and mathematics, but it also has major implications to real world. Its resolution could revolutionize the world. This article gives the definition and major results of P=NP.

Comments

  1. Great post! As a PhD student working in the same field, I find it difficult sometimes to explain to others what I’m doing. This is a great way to get people interested in the field. Good job!

  2. Good day! This is my first visit to your blog! We are a
    team of volunteers and starting a new initiative in a community
    in the same niche. Your blog provided us beneficial information to
    work on. You have done a extraordinary job!

Leave a Reply

Your email address will not be published. Required fields are marked *