Hello,
I was wondering if it is possible to specify up to which tolerance the equality constraints are satisfied in Optim.jl using IPNewton method.
Currently, the optimization stops because x_tol or f_tol are satisfied, but not because g_tol is satisfied. If I manually check the value of the constraints, they are satisfied within a tolerance of 1e-2. Is it possible to set a parameter to force higher precision?
I have tried setting x_tol=NaN and f_tol=NaN to force the algorithm to stop by g_tol, but then it stagnates:
Iter Lagrangian value Function value Gradient norm |==constr.| μ
0 3.332750e+01 1.746610e+00 3.438553e+01 1.901604e+01 1.00e+00
* time: 1.67178566274176e9
1 1.353145e+01 1.702241e+00 3.543331e+01 1.048336e+01 1.00e-01
* time: 1.671785662797074e9
2 4.377825e+00 1.699334e+00 7.754110e+02 3.833958e-01 9.76e-02
* time: 1.671785662809908e9
3 2.529626e+00 1.699267e+00 5.851341e+01 3.768007e-01 2.01e-02
* time: 1.671785662883332e9
4 2.278308e+00 1.703248e+00 8.222396e+00 1.410696e-01 1.99e-02
* time: 1.671785662899112e9
5 1.965445e+00 1.712061e+00 2.599936e+00 1.133042e-03 1.12e-02
* time: 1.67178566292044e9
6 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.67178566309099e9
7 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663162644e9
8 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663237643e9
9 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663344027e9
10 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663460111e9
11 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663539827e9
12 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663623636e9
13 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663723319e9
14 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663871442e9
15 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785663952723e9
16 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785664022078e9
17 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785664093651e9
18 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785664163002e9
19 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785664231979e9
20 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785664306802e9
21 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785664431646e9
22 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785664514846e9
23 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.671785664598036e9
24 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.67178566466456e9
25 1.766109e+00 1.712061e+00 4.019143e+00 1.133042e-03 2.35e-03
* time: 1.67178566474315e9
Why doesn’t it try to reduce the value of μ and see if it can improve the norm of the gradient?
I apologize for not posting the code (it is a research code and it has not been published yet).
Thank you very much,
SHCE