Question

altavir · Answer

It seems to be dead By the way I ve started porting my optimization codes to the multiplatform it includes Commons maths fitters a fork of JMinuint package and unique implementation of Fyodor Tkachev s QOW method I do not think it will be a part of Kmath since it is too complicated and relies on other my libraries Sadly I currently do not have time to work on it properly

altavir · Answer

But in order to make it multiplatform I will need eigenvalue decomposition

breandan · Answer

I ve heard is pretty fast portable ex JBlas recently switched to OpenBLAS without much trouble

breandan · Answer

Lately I ve been using some linear programming tools in my research I tried koptim and but they do not seem to scale very well so I switched to in the end At some point it would be nice to investigate porting some algorithms for LP MILP to KMath Apparently you can use the same power iteration technique that Alya showed for PCA on other semirings to solve many interesting problems in combinatorial optimization This video from Fred Hamprecht has a nice introduction and some applications of the idea

Ролан · Answer

I have experience with constrained optimisation working in finance My outcome is that if you can afford a commercial licenise you should go for Gurobi Otherwise in open source for continuous problems I recommend for mixed integer

Ролан · Answer

Regrading BLAS if on Intel machine you are better of with MKL than OpenBlas or Atlas if you have Nividia go for cuBLAS

Ролан · Answer

Regarding power methods and multiplatform implementation well yeah that would be awesome to see

breandan · Answer

Yeah someone else recommended Gurobi and apparently they provide an academic license so that also looks like a good option Another OSS alternative I looked at is MiniCP <http www minicp org|www minicp org>

Ролан · Answer

Yes Gurobi is the best optimiser by a mile for constrained problems especially for MILP The academic license should cover most of the functionality for free Please note as well that depending on your use case a powerful optimiser is not always enough You need also a good pre solver for the modelling language you use The best one I believe is AMPL again from personal experience However this is again commercial software but probably accessible for academics

Ролан · Answer

I am scared MiniCP wouldn t be able to handle anything beyond toy problems

breandan · Answer

Re Semirings I really like the way they re implemented in Torch Struct Implementing an algorithm like Viterbi with semiring algebras might be a good intern project I know Maria already has some code for this

Ролан · Answer

Looking at genbmm implementation I really find PyTorch to have grown up into an amazing computational platform The integration with Python C C++ amp CUDA is as seamless as it gets

Ролан · Answer

that s exactly what one needs writing numerical code

altavir · Answer

We ve just added boolean algebra and I thinkg we can add anything without breaking the compatibility so PRs are welcome