[petsc-users] Unconstrained optimization question

[Palmer, Bruce J]

I needed to add a call to TaoSetFromOptions to get the runtime options to work. I also set grtol from the default value of 1.0e-7 to 1.0e-15. The calculation goes for a few iterations, but it looks like it keeps pushing into territory where the objective function blows up. It eventually quits with a line search error. The complete output from both the runtime options and the TaoView command looks like


  4 TAO,  Function value: 9.24265e+85,  Residual: 3.11789e+76

  TAO  solve did not converge due to DIVERGED_LS_FAILURE iteration 4

Tao Object: 1 MPI process

  type: cg

    CG Type: prp

    Gradient steps: 0

    Reset steps: 4

  TaoLineSearch Object: 1 MPI process

    type: more-thuente

    maximum function evaluations=30

    tolerances: ftol=0.0001, rtol=1e-10, gtol=0.9

    total number of function evaluations=0

    total number of gradient evaluations=0

    total number of function/gradient evaluations=0

    Termination reason: -3

  convergence tolerances: gatol=1e-08,   steptol=0.,   gttol=0.

  Residual in Function/Gradient:=3.11789e+76

  Objective value=9.24265e+85

  total number of iterations=4,                          (max: 100)

  total number of function/gradient evaluations=20,      (max: 4000)

  Solver terminated: -6   Line Search Failure

I’ll have to bite the bullet and convert everything into ps, nm, and au to get values down to a range where they are all of order 1 to get this to work.

Bruce

From: Jed Brown <jed at jedbrown.org>
Date: Wednesday, June 26, 2024 at 3:02 PM
To: Palmer, Bruce J <Bruce.Palmer at pnnl.gov>, Barry Smith <bsmith at petsc.dev>
Cc: petsc-users at mcs.anl.gov <petsc-users at mcs.anl.gov>
Subject: Re: [petsc-users] Unconstrained optimization question
You can use the PETSC_OPTIONS environment variable to specify options if you don't pass the command line arguments through.

You can set -tao_grtol smaller to handle this difference in scales between the objective and the gradient, though applying some nondimensionalization/choice of appropriate units is still recommended.

"Palmer, Bruce J via petsc-users" <petsc-users at mcs.anl.gov> writes:

> This is a fortran code that doesn’t make use of argc,argv (I tried running with the runtime options anyway, in case you implemented some magic I’m not familiar with, but didn’t see anything new in the output). I have a call to TaoView(tao, PETSC_VIEWER_STDOUT_SELF,ierr) in the code and it reports back
>
>
>
> Tao Object: 1 MPI process
>
>   type: cg
>
>     CG Type: prp
>
>     Gradient steps: 0
>
>     Reset steps: 0
>
>   TaoLineSearch Object: 1 MPI process
>
>     type: more-thuente
>
>     maximum function evaluations=30
>
>     tolerances: ftol=0.0001, rtol=1e-10, gtol=0.9
>
>     total number of function evaluations=0
>
>     total number of gradient evaluations=0
>
>     total number of function/gradient evaluations=0
>
>     Termination reason: 0
>
>   convergence tolerances: gatol=1e-08,   steptol=0.,   gttol=0.
>
>   Residual in Function/Gradient:=7.54237e+75
>
>   Objective value=2.96082e+86
>
>   total number of iterations=0,                          (max: 100)
>
>   total number of function/gradient evaluations=1,      (max: 4000)
>
>   Solution converged:    ||g(X)||/|f(X)| <= grtol
>
>
>
> Bruce
>
> From: Barry Smith <bsmith at petsc.dev>
> Date: Wednesday, June 26, 2024 at 2:02 PM
> To: Palmer, Bruce J <Bruce.Palmer at pnnl.gov>
> Cc: petsc-users at mcs.anl.gov <petsc-users at mcs.anl.gov>
> Subject: Re: [petsc-users] Unconstrained optimization question
> Check twice before you click! This email originated from outside PNNL.
>
>
>   Please run with -tao_monitor -tao_converged_reason and see why it has stopped.
>
>   Barry
>
>
>
> On Jun 26, 2024, at 4:34 PM, Palmer, Bruce J via petsc-users <petsc-users at mcs.anl.gov> wrote:
>
> This Message Is From an External Sender
> This message came from outside your organization.
> Hi,
>
> I’m trying to do an unconstrained optimization on a molecular scale problem. Previously, I was looking at an artificial molecular problem where all parameters were of order 1 and so the objective function and variables were also in the range of 1 or at least within a few orders of magnitude of 1.
>
> More recently, I’ve been trying to apply this optimization to a real molecular system. Between Avogadro’s number (6.022e23) and Boltzmann’s constant (1.38e-16) combined with very small distances (1.0e-8 cm), etc. the objective function values and the values of the optimization variables have very large values (~1e86 and ~1e9, respectively). I’ve verified that the analytic gradients of the objective function that I’m calculating are correct by comparing them with numerical derivatives.
>
> I’ve tried using the LMVM and Conjugate Gradient optimizations, both of which worked previously, but I find that the optimization completes one objective function evaluation and then declares that the problem is converged and stops. I could find a set of units where everything is approximately 1 but I was hoping that there are some parameters I can set in the optimization that will get it moving again. Any suggestions?
>
> Bruce Palmer
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20240627/72b8e75c/attachment-0001.html>