Open anriseth opened 6 years ago
instead of quasi-newton how about a Newton-Krylov scheme? Should be easier to build into the current code.
instead of quasi-newton how about a Newton-Krylov scheme? Should be easier to build into the current code.
That would be interesting as well, sure.
instead of quasi-newton how about a Newton-Krylov scheme? Should be easier to build into the current code.
To clarify, do you mean supporting "Hessian-free" (Hessian-vector) type Newton-Krylov, or to support, for example, IterativeSolvers.jl
? I have only dealt with Newton-Krylov in the context of nonlinear equations, where it can also mean that you use some iterative solver for the problem (Jacobian*dx = Residual), regardless of whether the Jacobian is assembled or not.
In A NK scheme you can replace the hxv with finite differences- that is the advantage for me
I've added Newton-Krylov to the list :)
I will start updating the MathProgBase interface.
I will start updating the MathProgBase interface.
Thanks for doing that, @gragusa :) Note that we have added the IPNewton
code from this repository to Optim
now, so I think you should be able to update the interface with using Optim
instead of ConstrainedOptim
. Ref https://github.com/JuliaNLSolvers/Optim.jl/pull/608 https://github.com/JuliaNLSolvers/Optim.jl/issues/609
I also see that the JuliaOpt group have started implementing the solver interfaces with MathOptInterface now, as they are planning to move away from MathProgBase (Refs https://github.com/JuliaOpt/Ipopt.jl/pull/108 https://github.com/JuliaOpt/Ipopt.jl/pull/116). I don't know whether it is more complicated to interface to MOI and therefore less time consuming to do this for MPB, but it seems like MOI is the future.
Yes, SQP methods are certainly on the table. Notice that his functionality is now in Optim.jl :)
Here's a list of things that I believe we should focus on next. If anyone is interested in helping out here that would be very much appreciated.
Interface
value!
et al. for the constraints in NLSolversBase top priorityInterior / IPNewton
constraints.jl
againallow_f_increases
for constrained problems (it should "always" be allowed)Linesearch
backtrack_constrained
(onlybacktrack_constrained_grad
works at the moment)Testing