Posted
about 3 years
ago
by
vdebuen
Hello (Sorry for my poor english) I'm Victor de Buen, from Bayes Forecast  Spain, the main developer of open source project TOL, and I'm thinking to create a port package to METSlib to be used by TOL users in order to solve complex
... [More]
problems of real bussines world.
In this kind of problems, we can do almost no assumption over domain, restrictions nor behaviour of variables and cost function. Some times, evaluation of cost function is not analytical, you cannot express as a mathematical formulae, you have only a blackbox process that takes a set of parameters and returns a value.
For continuous domain problems we use IPOPT, when gradient and hessian are available, and NLOPT when we have only the gradient or even a non continuous cost function.
We have no speciallized method for integer or mixedinteger problems, and I think that metaheuristic algorithms are the best option in real world problems.
I supose that algorithms will be much more efficient when user gives specific methods to manage moving, and other algorithm related methods, but I'm afraid that many times user don't know almost anything about the problem except the cost function. However, the CPU time is everytime cheaper, so it could be sufficient having a low convergence rating.
I've readed the documentation, the source code and the examples and I believe that is a good option for TOL due to it's written in C and have a wide battery of algorithms. I don't know much about metaheuristic methods and I don't understand very well the neighbourhoodrelated methods. It's mandatory to define a problemdependent moving methods, or it exists a default one that works for a blackbox cost function? I have the same doubt about specific methods to configuration of algorithms like tabu related methods.
Thanks in advance. [Less]
