Open-source is an approach to “publishing” software -- (1) peer-review (2) archived

47 out of 125 coin-discuss subscribers are from company email addresses (this excludes “.com” from yahoo, etc.)

International representation (US, Brazil, Canada, Germany, Italy, Switzerland, …)

Individuals now at Hotchkiss, SAS, US Air force, and Research and Development Division, Electricity de France have contributed to COIN-OR.  University to companies (Examples: Matthew Galatia and Yan Zu worked with COIN-OR at Lehigh U. Now at SAS. Andreas W: CMU to IBM.)

Stefano Gliozzi, Managing Consultant IBM Business Consulting Services.  stefano_gliozzi @ it.ibm.com (1) We are using it on a large Yield Management project for a Railway.  COIN/OSI/CLP Both in the prototyping phase and the production code. (2) We are doing a preliminary study (IBM Internal, yet not signed by the client ) for a Space Allocation / Static and Dynamic balancing of payloads for the European launcher of the ISS. This would be version 2 of an application developed using IBM OSL.. we are using COIN/OSI/CLP/CBC/CGL (3)  Finally as an IBM Asset , we developed a version of EasyModeler, a translator between an Algebraic Language and C, which produces code able to interface the COIN/OSI. Presently I tested it using COIN/OSI/CLP, COIN/OSI/CBC/CGL, COIN/OSI/OSL

Hosted by INFORMS in a “fee for service” arrangement - $1/year – to avoid legal liability.

Projects include (1) frameworks, (2) interfaces, (3) solvers, and (4) “other” utilities and functions.

Fabrizio Ferrandi, Dipartimento di Elettronica e Informazione, Politecnico di Milano. 2 uses. (1) We are currently using CBC in our research aiming at high-level synthesis of C --------------------------------------------------------------------------------------------------------------- Abstract: Recent research results have seen the application of parallelizing techniques to high-level synthesis. In particular, the effect of speculative code transformation on mixed control-data flow designs has demonstrated effective results on schedule  lengths. In this paper we first analyze the use of control and data dependence graph as  an intermediate representation that provides the possibility of extracting the maximum parallelism. Then we analyze the scheduling problem by formulating an approach based on Integer Linear Programming (ILP) to minimize the number of control steps given the amount of resources. We improve the already proposed ILP scheduling approaches by introducing a new conditional resource sharing constraint which is then extended to the case of speculative computation. To efficiently solve the ILP formulation we use a standard Branch and Cut framework. Experimental results show that this framework gives better results ------------------------------------------------------------------------------------------------------------------------- (2) Moreover, we are thinking to port on the COIN-OR framework an old work on logic synthesis based on Lagrangean relaxation. This work has been published and this is the reference:  R. Cordone, F.Ferrandi, D. Sciuto, R. Wolfler Calvo, “An Efficient Heuristic Approach to Solve the Unate Covering Problem”  IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 20, n. 12, December 2001, pp. 1377 – 1388.  In case of interest this is the abstract: --------------------------------------------------------------------------------------------------------------- Abstract: The paper presents a new approach to solve the unate covering problem based on  exploitation of information provided by  agrangean relaxation. In particular,  main advantages of the proposed heuristic algorithm are the effective choice of elements to be included in the solution, cost-related reductions of the problem, and a good lower bound on the optimum. The results support the effectiveness of this approach: on a wide set of benchmark problems, the algorithm nearly always hits the optimum and in most cases proves it to be such. On the problems whose optimum is actually unknown, the best known result is strongly improved ------------------------------------------------------------------------------------------------------------------------- Yes, we are using CLP/CBC in our product. Our product is still pre-release and therefore we cannot discuss it in any detail at the moment unfortunately. We expect to make a public release in a matter of months. We have found Clp/Cbc to be a fantastic toolset for LP and MIP problems. It is very robust and quite mature, though there are clearly significant and very desirable improvements planned for the future and under development. We use a number of features of CBC which are not universally supported by other solvers and will over time use the tuning features of CBC to improve performance further. In particular we use generalised lot-sizing (more than just semi-continuous variables) and The API has been on the whole easy to work with given that we have used it as it has developed. John Forrest has been exemplary in responding to our questions and troubleshooting and we would particularly like to express our gratitude to him. Our product is designed to be able to support a number of different LP/MIP codes as plug in modules. This is important to us so that we can compare and contrast different codes and pick the best for our needs. At the moment we choose Clp/Cbc as the best for these needs. We will be going through another evaluation cycle before final product release, so this may change of course. I hope the above is of some use to you, in particular we would like to say that having such a high quality MIP code available for use by startup companies such as our venture is a wonderful enabling technology and I am sure many will follow this path in the future. Here is some more information about our work with COIN here at GAMS. Hopefully it helps to prepare your talk for Monday. About two years ago, GAMS started to build a small object layer to connect the COIN/OSI interface and our GAMS IO library so we could feed OSI with LP and MIP problems generated by GAMS. This small project turned out to be a highly successful. One the one hand we could offer our clients reliable access to open source mathematical programming solvers like Cbc and Glpk. On the other hand with the help of our clients and our rigorous testing facilities, we have been finding and fixing bugs in COIN that hopefully improved the overall quality of COIN. We have about 100 downloads of the GAMS system a day. Each and everyone contains the COIN solvers. We unfortunately don't have the sales staff to know exactly what our client are doing, but we frequently hear about people working with the COIN solvers (sometime with a description of a bug). For example, in Germany we have a company (I am not allowed to reveal their name) using Cbc on a daily bases to solve medium sized periodic scheduling problems. Moreover, Josef Kallrath created a cutting stock application written in GAMS used in several paper mills that in its core solves several small to medium sized MIPs with Cbc. We hope that we can intensify the work between GAMS and COIN, e.g. by contributing to a standard way of interfacing with non-linear programming solvers, an area where GAMS has acquired expertise for the last two decades.

 - Cincinnati Children's Hospital Research Foundation's Biomedical  Informatics division.  They have been using SYMPHONY for solving some bioinformatics problems. What is the relationship of SYMPHONY and COIN/BCP?
COIN/BCP is a C++ framework based roughly on an ancestor of SYMPHONY, so they take substantially the same approach and have similar functionality. COIN/BCP was developed at IBM and is currently part of the COIN-OR software repository. Currently, SYMPHONY is easier to use "out of the box" and is a better choice for beginners. It is also more efficient for performing pure branch and cut (without column generation). COIN/BCP has improved on SYMPHONY in some areas, however, such as support for branch and price. SYMPHONY's support for branch and price is limited. COIN/BCP is missing some features that SYMPHONY has, such as an implementation of global cut pools. SYMPHONY and COIN/BCP may be combined into a third-generation framework currently under development.

Brady Hunsaker, U of Pgh: Together with a collaborator at Georgia Tech, I use OSI for an NSF-funded (DMII OR) research project on using software to automate the search for valid inequalities of low Chvatal-Gomory rank.  We are creating an open-source tool that will be usable with any OSI solver.

2 interfaces – (1) IPOPT (2) Lancelot.

SMI generates the deterministic equivalent of a multiperiod stochastic program for any OSI compatible solver.  It implements the scenario and independent discrete distribution formats of the Stochastic MPS standard. It is being used primarily by academics building models to research properties of stochastic programs.

DFO is a Fortran package for solving general nonlinear optimization problems that have the following characteristics: their objective function is relatively expensive to compute and derivatives of such functions are not available and There also may be some noise in the function evaluation procedures. Such optimization problems arise ,for example, in engineering design, where the objective function evaluation is a simulation package treated as a black box. Well, it is a code that targets general nonlinear optimization  problems with expensive function evaluations and without derivative computation. This is one of the few (so far) state of the art codes. It is very easy to use (since practically all the user needs is to write a routine that computes a function value). It is also Another unique feature is that it handles various kinds of constraints, such as derivative free, or general nonlinear 2. There is a fair number of users, but what exactly they are doing we do not know, usually. There is a recent paper by Shafer, Karasozen, Uludag, Yapici, Ogur from Middle Eastern Technichal University in Ankara, Turkey on optimization of stirrers (for chemical, biotechnological, pharmaceutical and food processes). They use DFO as their optimization engine. The problems fall into the category that DFO targets. DFO is also used with DEFT - Derivative Free Tuner, a circuit tuning tool that was developed at IBM for tuning of small scale circuits which require high accuracy provided by PowerSpice simulator. This simulator does not compute or approximate derivatives. Deft is the first tool based on mathematical sound optimization tool (DFO) and it works much better than the predecessor. It is OK to talk about this publicly. Specifically DFO was used successfully for tuning clock distribution networks

dylp is designed to be the lp module in an lp-based branch-cut-price mixed-integer linear programming code. Very efficient when reoptimising after small modifications to the problem. (The active system changes slowly and remains minimal.) The dynamic simplex algorithm is also well suited for column generation algorithms. dylp is designed as a research code, and can provide copious trace output and extensive consistency checks. It's fast with an optimized build, and downright paranoid when all consistency checks are enabled. A design goal was that it should never die before telling you exactly what went wrong, and where. Can be forced to use the full constraint system and behave like a normal simplex algorithm. The dynamic simplex algorithm is laid out in Padberg, ``Linear Optimisation and Extensions'' (1991), and likely in lots of other places. The basic idea is to always work with a partial constraint system and dynamically adjust the active system so that it contains only those constraints and variables that are needed for the current extreme point and immediate surroundings. The core of the implementation is two alternating cycles: purge/add variables and reoptimize with primal simplex;  purge/add constraints and  reoptimise with dual simplex.

POOPT is the only interior-point-nlp-solver that’s both open-source and “good”. IPOPT is the only one  that uses a “globally convergent” filter algorithm. (Distinguishes it from Loco). Pennon (Germany) is another interior point solver – they solve *more* general problems but at this point performance lower on IPOPT’s class of problem .  2.  Several research projects in Larry Biegler's group at CMU, including one with Ipopt as optimization engine in the process simulator ROMeoTM (slides)
3. In the IBM-internal circuit tuning tool EinsTuner 4. In ABB's Pulp and Paper Online Production Optimizer
5. Tobias Jockenhövel's OptControlCentre

6. By Ned Nedialkov in the context of solving Differential-Algebraic Equations

7. At the Munich University of Applied Sciences (for the course on "Modeling, Simulation and Optimization" by Dieter Kraft) 8. By Juan Arrieta, in the computation of optimal trajectories for spacecraft and related astronautics problems

- The Volume Algorithm has been very effective at producing approximate solutions to large - A search in Google Scholar found 64 articles that cite the VA. This is relatively high for an article published in 2000. - At IBM the VA was used for (1) set partition problems coming from airlines optimization and (2) later for

>You can find some CppAD "graphics" under the Theory section of the manual. > http://www.seanet.com/~bradbell/CppAD/theory.xml > http://www.seanet.com/~bradbell/CppAD/theoryrevnotation.xml > http://www.seanet.com/~bradbell/CppAD/reverseidentity.xml

1. The main groups using multifario are Sandia, where they are looking at parameterized circuits (they have something called Xyce, which is like SPICE). They helped do the interface between LOCA and multifario, but it sits in the LOCA code, so I don't know if that counts. 2.The other group is John Maddocks at EPFL in Lausanne Switzerland who does DNA folding. A former postdoc of his, Sebastien Neukirch used multifario to study a uniform elastic rod, and John has plans to use EPFL's new BlueGene for more complicated models and we have plans to parallelize multifario for that.” - MH

Logo? That’s a beaver! Mascot of Oregon State University.