L-BFGS-B optimizer appears to fail

[gameoverhack]

Many thanks for porting this to Tensorflow.

Everything seems to work for me if I use the Adam optimizer, however when I try LBFGS I get output that is not by stylized. It appears that the optimizer stops after only one iteration, emitting various warnings (randomly it seems).

My machine has 64Gb Ram and a GTX 1080 Ti.

Here are a few outputs when I try LBFGS...any help would be greatly appreciated.

Machine precision = 2.220D-16
 N =       874800     M =           10
 This problem is unconstrained.

At X0         0 variables are exactly at the bounds

At iterate    0    f=  2.88465D+12    |proj g|=  1.09390D+07

           * * *

Tit   = total number of iterations
Tnf   = total number of function evaluations
Tnint = total number of segments explored during Cauchy searches
Skip  = number of BFGS updates skipped
Nact  = number of active bounds at final generalized Cauchy point
Projg = norm of the final projected gradient
F     = final function value

           * * *

   N    Tit     Tnf  Tnint  Skip  Nact     Projg        F
*****      2     18      1     0     0   7.054D+06   2.124D+12
  F =   2123653447680.00     

CONVERGENCE: REL_REDUCTION_OF_F_<=_FACTR*EPSMCH             

 Warning:  more than 10 function and gradient
   evaluations in the last line search.  Termination
   may possibly be caused by a bad search direction.

 Cauchy                time 0.000E+00 seconds.
 Subspace minimization time 0.000E+00 seconds.
 Line search           time 0.000E+00 seconds.

 Total User time 0.000E+00 seconds.

==========================================

Machine precision = 2.220D-16
 N =       874800     M =           10
 This problem is unconstrained.

At X0         0 variables are exactly at the bounds

At iterate    0    f=  2.88465D+12    |proj g|=  1.09390D+07

 Nonpositive definiteness in Cholesky factorization in formt;
   refresh the lbfgs memory and restart the iteration.
  ys= 0.000E+00  -gs= 0.000E+00 BFGS update SKIPPED

           * * *

Tit   = total number of iterations
Tnf   = total number of function evaluations
Tnint = total number of segments explored during Cauchy searches
Skip  = number of BFGS updates skipped
Nact  = number of active bounds at final generalized Cauchy point
Projg = norm of the final projected gradient
F     = final function value

           * * *

   N    Tit     Tnf  Tnint  Skip  Nact     Projg        F
*****      6     15      3     1     0   6.814D+06   2.102D+12
  F =   2102054223872.00     

CONVERGENCE: REL_REDUCTION_OF_F_<=_FACTR*EPSMCH             

 Cauchy                time 0.000E+00 seconds.
 Subspace minimization time 0.000E+00 seconds.
 Line search           time 0.000E+00 seconds.

 Total User time 0.000E+00 seconds.

Frame 1 elapsed time: 10.81877875328064

==========================================

Machine precision = 2.220D-16
 N =       874800     M =           10
 This problem is unconstrained.

At X0         0 variables are exactly at the bounds

At iterate    0    f=  2.88465D+12    |proj g|=  1.09390D+07
  ascent direction in projection gd =    13099410.6106499     

 Bad direction in the line search;
   refresh the lbfgs memory and restart the iteration.

           * * *

Tit   = total number of iterations
Tnf   = total number of function evaluations
Tnint = total number of segments explored during Cauchy searches
Skip  = number of BFGS updates skipped
Nact  = number of active bounds at final generalized Cauchy point
Projg = norm of the final projected gradient
F     = final function value

           * * *

   N    Tit     Tnf  Tnint  Skip  Nact     Projg        F
*****      2      9      2     0     0   6.981D+06   2.122D+12
  F =   2122045587456.00     

CONVERGENCE: REL_REDUCTION_OF_F_<=_FACTR*EPSMCH             

 Cauchy                time 0.000E+00 seconds.
 Subspace minimization time 0.000E+00 seconds.
 Line search           time 0.000E+00 seconds.

 Total User time 0.000E+00 seconds.

Frame 1 elapsed time: 10.421173810958862

[gloderndelut]

Have the same issue. Did you find a solution?

[Ashwanthkumard]

Facing the same issue while running the ARIMA models. Any solution to overcome?

[ne-bo]

Same here.

[Billy1900]

Same here, seems like there are few people who could meet this problem.