Skip to content

Latest commit

 

History

History
72 lines (50 loc) · 2.29 KB

to_do.md

File metadata and controls

72 lines (50 loc) · 2.29 KB
Raw
  • low tolerance Ipopt
  • infeasible Ipopt

Short term:

  • mu autotune scaling -- run on whole test set.

  • increase delta to min stable line search??? test

  • new delta strategy with loglog rate???

  • write code that saves all the CUTEst files need for the paper in one go! (i.e., have an experimental script and a paper producing script)

  • add COPS problems

  • improved initial delta selection

Medium term:

  • detect is problem is caused by factorization issues or lack of smoothness of functions (identify function, a direction and a point)

  • improve efficiency of schur complement and eval_jac

  • deal correctly with exceptions

  • add parameters for termination criterion

  • test unconstrained problems

- create infeasible test set - scaled termination criterion - run full netlib test

*- mu stuff *

- initialization scheme

  • fix error with predicted reduction of complementarity

  • stabilization should prioritize complementarity if dual is small and comp not satisfied

  • measure complementarity in output in relative terms

  • split up parameters i.e. initialization parameters etc ...

  • install IPOPT on sherlock

  • KKT system S_diag and X_diag as variables

  • increase delta when ever there is any sort of failure

  • add option to move dual and primal iterates independently

  • mu choice?

  • add protections to aggressive steps

  • move dual slowly? maybe not i think one is better off regularizing. #- corrections, only do if predicted progress is good (and stop line search immediately if it is not).

  • solve MUMPS issues, version #- add symmetric KKT system solver and deal with inaccuracy #- write proper line search for stabilization step #- non-linear updates of primal variables

  • infeasiblity detection

  • unboundedness detection

  • momentum/homogenous style scaling

  • create notes of what I am doing

  • validate LP direction

Long term:

  • (1) find LP solution first, (2) start from analytic centre, (3) re-write so problem is well-conditioned

  • automatic scaling trust region algorithm

  • momentum/CG in stabilization steps

  • re-use permutations for cholesky

  • filter during stable steps, either accept improvement in dual or primal

Blue sky

  • run non-linear CG to 0.99 accuracy in 100 iterations. Use NC certificates to deduce correct delta.
  • do linear algebra to deal with duplicates of constraints