[ADMB Users] Bug? Limited memory quasi newton terminating with the wrong gradient

[lennart at ii.uib.no]

Hi,

I have been doing some simple testing of ADMB to see if I could get limited memory quasi-newton to work. However, on one problem it reports the initial solution as the final solution and the gradient as zero, both of which are wrong. If I don't use -lmn, then everytying is fine. I'm wondering if this is just a problem with my installation, or if there is a bug somewhere in ADMB.

In more detail:

ADMB version: 9.1 (December 31, 2009)
Architecture: 64 bit, four processors of type Intel(R) Xeon(R) CPU  X3330  @ 2.66GHz
OS: Linux
Source code compiled with either make linux or make linux-64 (to the best of my recollection probably the first one)


The function is the Rosenbrock function (this one http://en.wikipedia.org/wiki/Rosenbrock_function ). If I compile and run with the standard optimization method then I get:

./rosenbrock

Initial statistics: 2 variables; iteration 0; function evaluation 0
Function value   2.4200000e+01; maximum gradient component mag  -2.1560e+02
Var   Value    Gradient   |Var   Value    Gradient   |Var   Value    Gradient   
  1 -1.20000 -2.15600e+02 |  2  1.00000 -8.80000e+01 |

Intermediate statistics: 2 variables; iteration 10; function evaluation 17
Function value   2.4618636e+00; maximum gradient component mag  -1.0410e+01
Var   Value    Gradient   |Var   Value    Gradient   |Var   Value    Gradient   
  1 -0.53029 -1.04105e+01 |  2  0.24656 -6.93000e+00 |
2 variables; iteration 20; function evaluation 29
Function value   1.9826294e-01; maximum gradient component mag  -3.0528e+00
Var   Value    Gradient   |Var   Value    Gradient   |Var   Value    Gradient   
  1  0.58171  2.71511e+00 |  2  0.32313 -3.05278e+00 |
2 variables; iteration 30; function evaluation 42
Function value   1.0044269e-03; maximum gradient component mag   7.8688e-01
Var   Value    Gradient   |Var   Value    Gradient   |Var   Value    Gradient   
  1  0.97657  7.86878e-01 |  2  0.95156 -4.26868e-01 |

 - final statistics:
2 variables; iteration 36; function evaluation 48
Function value   4.0456e-13; maximum gradient component mag   9.5059e-06
Exit code = 1;  converg criter   1.0000e-04
Var   Value    Gradient   |Var   Value    Gradient   |Var   Value    Gradient   
  1  1.00000  9.50595e-06 |  2  1.00000 -5.33050e-06 |
Estimating row 1 out of 2 for hessian
Estimating row 2 out of 2 for hessian

...which is the right solution. If, however, I use limited memory quasi-newton, then the following happens:

./rosenbrock -lmn 5

Initial statistics: 2 variables; iteration 0; function evaluation 1
Function value   2.4200e+01; maximum gradient component mag   2.1560e+02
Var   Value    Gradient   |Var   Value    Gradient   |Var   Value    Gradient   
  1 -1.20000 -2.15600e+02 |  2  1.00000 -8.80000e+01 |

final statistics: 2 variables; iteration 1; function evaluation 2
Function value   1.7134e+02; maximum gradient component mag   2.6055e+02
Var   Value    Gradient   |Var   Value    Gradient   |Var   Value    Gradient   
  1 -0.27415  1.40311e+02 |  2  1.37790  2.60547e+02 |
Estimating row 1 out of 2 for hessian
Estimating row 2 out of 2 for hessian


Which is premature termination. That in itself does not necessarily imply a bug, but if I look at the .par-file, it reads:

# Number of parameters = 2  Objective function value = 24.2000  Maximum gradient component = 0.00000
# x1:
-1.20000000000
# x2:
1.00000000000

In other words, it reports the initial solution as the final solution,
and the corresponding gradient as identically zero, both of which are clearly wrong.


The code etc is:

rosenbrock.tpl:

DATA_SECTION

  init_number const_1;
  init_number const_2;


PARAMETER_SECTION

  init_number x1;
  init_number x2;

  vector f_vec(1,2); 
  objective_function_value f;

PRELIMINARY_CALCS_SECTION

PROCEDURE_SECTION

  f_vec(1) = const_1*(x2 -x1*x1);
  f_vec(2) = const_2*(1-x1);
  f =  norm2(f_vec);



REPORT_SECTION

TOP_OF_MAIN_SECTION


rosenbrock.dat: 

# Parameters for rosenbrock function, that is, constants
10 1

rosenbrock.pin:

# Starting values for rosenbrock
-1.2 1


Best regards,

Lennart