The Java/Scala choice to implement CP Solvers

I recently started to implement a solver in Java during my free time. I choosed Java for several reasons:

- it is still by far the most popular language

- integration and deployment facility of a Java project + portability

- most of my colleagues at n-Side know Java and they are less familiar with C++

- the development tools (eclipse/InteliJ) are wonderful

- good tradeoff between speed and ease of modeling

Of course the price to pay is this factor 3. I like a lot the point of view of Nate Brixius (developer of Solver Foundation) about that. Here are few extracts about his blog post:

"I did standard critical path scheduling in both C# and C++ and found that the C# version was only 10-15% slower. We went with C# and in the course of development we were able to find algorithmic improvements that resulted in a faster implementation that the original C++ codebase. "

"Many programmers (myself included) are much more productive writing C# code. This gives me more time to investigate possible algorithmic improvements, write good unit tests, etc. There are opportunity costs!"

This is also true for CP solvers, what really makes a difference to solve a problem with CP is the functionalities available in your solver: the modeling facility, the global constraints available, how easy it is to write a custom search, to do a LNS, restarts, make parallelisation ...

I'm certainly not the first one to have chosen Java to develop a CP solver. Two popular open source CP solvers are implemented in Java: Choco and Jacop. There is also a standardization project for solvers in Java.

One of my colleague (Gilles) at n-Side is a real "Scala Enthutiast" and made me discover this wonderful language called Scala.

Since that time, I don't think anymore that writing a model in Java is very convenient...

The really good thing about Scala is that its syntax is so flexible that you can use it to write DSL (If find Scala much better than Python to do that)

But the other cool thing about Scala is that it is at least as efficient as Java.

A comparison between C++, Java, Scala and Go made by Google was published at the Scala Days a few days ago on an algorithmic benchmark:

Scala was about 3 times slower and Java 5 times slower than an optimized C++ code.

Of course I didn't throw away all my Java implem but I started a Scala wrapper on top of it for fun (integration between Scala and Java is perfect in both direction). I choosed to make a syntax quite close to OPL/Comet because I'm used to it and I find it very clear. Here is how the Queen problem looks like in my Scala wrapper:

object Queens  extends Model {
   def main(args: Array[String]) {
  
      val cp = new CPSolver()
      
      val n = 6 //number of queens
      val Queens = 0 until n
      //variables
      val queens = for(i <- Queens) yield CPVarInt(cp,1 to n)
      
      var nbsol = 0
      cp.onSolution {
       nbsol += 1
      }
      
      cp.solveAll subjectTo {
       cp.add(alldifferent(queens))
       cp.add(alldifferent(for(i <- Queens) yield queens(i) + i))
       cp.add(alldifferent(for(i <- Queens) yield queens(i) - i))
      } exploring {
       //exploration of the search tree
       new Binary(queens)
      }

      //print some statistics
      println("#sol",nbsol)
      cp.printStats()
    }
}

Again, it seems that I'm not the only one to have taken that decision. The Jacop team also adopted the same approach and started to make a wrapper for their solver. Here is the Queen model in Jacop/Scala:

object Queen extends Application with jacop {

  val n = 50

  val q: List[IntVar] = for (i <- List.range(0, n)) yield new IntVar("q"+i, 0, n)

  def noattack(i: Int, j: Int, qi: IntVar, qj: IntVar) = {
 qi     != qj 
 qi + i != qj + j 
 qi - i != qj - j 
  }

  for (i <- 0 until n; j <- i+1 until n) noattack(i, j, q(i), q(j))

  val result = satisfy( search(q, first_fail, indomain_middle) )

  if (result) 
    q.foreach(qi => {
      for( i <- 0 until n) 
  if (qi.value() == i) print(" # ") else print(" . ")
      println

    })
  else println("No solution")
}

I think we may see even more Scala wrapper for CP solvers in the near future...

Cloud and iCloud

I am currently more and more involved in cloud computing management. Mainly because I am member of the Aeolus ANR project (see http://www.aeolus-project.org). The university for which I work, has also just hired Fabien Hermenier who comes from Ecole des Mines de Nantes and who was one of the developer of the Entropy system. Fabien has some very interesting problems of cloud management and he modeled some of them by using CP.  I hope to show that CP is a really efficient technique for solving these problems, mainly because it can easily integrate new additional constraints.

At the same time, Steve Jobs presented iCloud. I think this is really a revolution in cloud computing or maybe a revolution in marketing. Here is his idea: if you have musics on your computer that you bought from iTunes, now you will be able to access to this music from the iTunes music store! Great you still have the right to use what you paid for. But the true revolution of iCloud is the following: if you have some musics on your computer that you didn't buy from itunes (for instance you ripped a CD) then you will be able to access to this music from the iTunes music store for only $25 a year! They call this iCloud and as you can see on this slide

you have no longer to wait for the long upload like with Google or Amazon! In fact, they just need to upload the MP3 Tag! That's a new way to see cloud computing. Apple need just to know your MP3 Tags and upload them : how many lines of code for doing that ? 100 ? That's a revolution.
At the same they avoid a lot of problem like memory consumption, time transfer and so on... Note that they still need minutes for your MP3 tag.

In addition you buy one more time your music...

For sure, iCloud is different...

Blind reviewing

Here is a review I just received:
"The paper should be resubmitted with more references to the existing work about the All-Different constraint"

Open Source and Originality

Last week, a lot of people (including Pierre, Thierry and myself) attended the CP-AI-OR Conference in Berlin.

We had some discussions about a paper implementing Goal programming in an open source solver. I wondered what was the scientific contribution and I asked the question. In fact, the authors were not really able to answer to my question. It was a presentation associated with an abstract, thus I think that it is not really important. 

I am more interested in one argument that has been answered: "there is a scientific contribution because this is new for open-source". We discussed about this point. Here are some remarks. I am curious to know if you have some others:

First, it means that all existing "proprietary" stuff has absolutely no value because it is not open-source! This is a strange conception of the originality. I also wonder why the open-source has several licences which forbid to reproduce the open source concept or code (because rewritten is not equivalent to originality)... 

Then, some people suggested that the minimum of honesty  should be to present the ideas in that way:

"Hello all, I am going to present an implementation in my open source solver of some works that has been carried out in several non open source solvers. In addition, you need to know that these ideas was clearly documented in the other products. Therefore, our job was mainly to  integrate these works in our open source product"

I think it is an interesting point of view. However, I guess that it will be more complex to have a paper accepted with such an introduction.

At last, note that it is not mandatory to publish papers...

CP arrives in business schools!

One difficulty faced by the new CP users is that you need programming skills to use it because all solvers’ API can only be used from programming languages. This is probably the reason why in business schools, the simplex is used almost exclusively to illustrate optimization (you can solve linear problems in Excel!).

But this will change since AIMMS (with the help of Willem-Jan van Hoeve) has decided to add a CP modeling API as well (paper available here). I think the version is still beta and they are looking for people to test it on AIMMS' google group.

I believe it is very good to disseminate this great technology to make it available from high level modeling tools like AIMMS (only a few clicks and few numbers to encode and your CP model is ready). Now the professors in business schools have no excuse any more for not teaching CP ;-)

Actually this may only be the beginning of something: Robert Fourer (creator of the very popular AMPL) was there at the CPAIOR conference that just finished today. Maybe I speculate but we may see CP coming soon in AMPL as well. After all this plan of extension for AMPL for CP is there since a long time. Robert Fourer answered about this extension on AMPL mailing list last month "We have hopes that this long-delayed project will move forward in the next year or so."

As you see, only good news for CP because if AMPL and AIMMS offers CP, it will reach many people from the OR-INFORMS community.

The future will tell us if those dreams come true and CP will have the popularity it deserve… In the meanwhile we can only encourage AMPL team to build this CP extension.

Informs Analytics, some feedback

I’m just back from Informs Analytics. I was there with 4 of my colleagues from n-Side. We were presenting posters about our products (this is a picture of my poster)

It was really a nice experience and a nice organization. If you feel sometimes alone or isolated doing optimization… Go there ! You’ll see that analytics (the new buzzword for OR) is used in every big organizations.

It was also fun to put a face on names that you know from Internet (like Natan Brixius having a blog on Microsoft Solver Foundation, Bob Fourer from Ampl and many others from Or-Exchange).

I was also very seduced by the presentations on Knitro solver (non linear continuous optimization) made by Richard Waltz. Continuous optimization is a technology that we use a lot at n-Side and we learned a lot there thanks to Richard!

I also attended a presentation given by Greg Glockner (Gurobi) on how to debug/detect numerical issues of a MIP. He is a great presenter! I didn’t know for instance that if you write in a MIP a constraint like y <= 10000000000 x with x a binary variable, you can get into troubles and have y taking positive values even without forcing x to 1 due to rounding errors. I found it a bit fun. CP doesn’t have those kinds of numeric issues because it is discrete by nature.

Of course the battle of MIP solver was more fierce than ever (FICO, CPLEX, GUROBI, …). It seems that they are fighting for some percents of speed-up on some benchmarks (mittleman bench among others). Honestly, I’m not sure it really matters to the end user. First, we even don’t know what kind of problems is present in those benchmarks. Second, what I would like to see is how easy it is to make a column generation, a branch and price or a new cut…

The Edelman award was really impressive. A bit like a Hollywood show http://www.flickr.com/photos/informs/5611418031/in/photostream The winner was MIDWEST ISO. They converted a local electricity market to a larger regional one resulting in lower electricity costs and more efficient market. Of course all that taking the constraints of the network into account. That was really great to me and my colleague to see this winner because we have a very similar product at n-Side where we are in charge of coupling the European market.

My main regret was that CP doesn’t have the status it should have at INFORMS. The only place where I’ve seen a CP application was on the Monsanto poster of my friend Ravichandran Venkatesh. They use CP to solve a kind of hybrid placement/scheduling problem for greenhouses.

The reason may be that CP is too focused on the AI community and less on Informs. This is sad because many of the problems solved there could be solved by CP or by hybridization with CP. For instance I’ve seen a presentation given by Fico where they try to build optimal planograms for supermarkets. A planogram is something like that:

They use a column generation for that. Great! But they have a log of sequence like constraints (regular, stretch…). So I’ve told them that CP would be very appropriate to do that.

In conclusion, It would be nice to see in the future more CP there. For instance I would enjoy seeing Google winning the Edelman award solving a real application with his new CP Solver ;-).

Yes we can … optimize!

Recently, I heard from someone (that I was trying to convince how CP is cool) : « yes but you know, most of the time we have an objective function … ». This person has a very strong MIP background and it’s not the first time I heard such a belief. Needless to say this is wrong and no, CP is not only for satisfaction problems.

CP usually solves optimization problems with B&B by optimizing one variable having its own domain. In short, this is how it works:

When a solution is found, a constraint is dynamically added (without restarting the search) such that the next solution found is strictly better. The last solution found is by construction optimal (correct me if I’m wrong but I think this adaptation of B&B to CP was imagined by Pascal Van Hentenryck).

This is also completely false to say that CP doesn’t use lower bounds. Usually the objective function/variable is linked into one (or several) global constraint computing very good lower bounds on it (look for instance the paper of Jean-Charles: « Arc consistency for global cardinality constraints with costs ». With this kind of constraints you can solve TSP problems with CP. This is a small demo on google map I made some time ago http://travellingsalesmanproblem.appspot.com/ It is a bit slow due to the time needed to collect the distance matrix L