RGA - algorithm for black box global optimization


RGA  in near-real style.


 Inverse Helmholtz.

"Multibanana function"

The function  is induced by the inverse Helmholtz,  simulating 2.5 D microopical tomography
(acquisition is of cylindrical symmetry, cylindrical objects are illluminated by polarized monochrome linear  sources).


Inverse problem:
   given:    find  4  parameters: Objective functionL2-norm of residuals (differences of measured and synthetic data).


Fig. 8.   2D-slice  of multibanana function is displayed (two parameters of the object location are fixed) .  Global minimum corresponds to  6.00 -> radius of the object  and 1.25 -> refractive index value : marked with red rhombus
a)   refractive index range  is [  0.7 : 1.3 ]
b)   refractive index range  is [ 1.1  : 1.5 ].
2D-section of Multibanana functiona)
Zoom of a)b)


 The results of RGA-runs.

 The result  of RGA-run in a following domain of search
 
 y 
radius 
refractive index
[ 9.0 : 11.0 ] 
[ 6.0 : 8.0 ] 
[ 4.0 : 7.0 ] 
[ 0.7 : 1.4 ]

is as follows - with true values respectively
 
10.0007 
 6.9908 
6.0027 
 1.2495
10.
7.
6. 
1.25



Respective OF-value is 0.0001278.
The total number of OF-calls is 1486.



The result  of the other RGA-run:  another domain of search/true parameters
 
 y 
radius 
refractive index
[ -2.0 : 0.0 ] 
[ 5.0 : 7.0 ] 
[ 6.0 : 9.0 ] 
[ 0.7 : 1.5 ]

 is as follows:
 
-0.9719 
 6.0740 
8.0081 
 1.3995
-1.
6.
8. 
1.40



Respective OF-value is 0.0004004.
The total number of OF-calls is 1850.



 
'Corrupted' multibanana

Fig. 9.  2D-slice  of  the 'corrupted' multibanana function:  induced by the same  problem as it
     is represented with  Fig.8,   except the input data are just amplitudes (no  info  on phase).


 The results of RGA-runs.


The result  of RGA-run:  another domain of search/true parameters
 
 y 
radius 
refractive index
[ 9.0 : 11.0 ] 
[ 6.0 : 8.0 ] 
[ 4.0 : 7.0 ] 
[ 0.7 : 1.3 ]

 is as follows:
 
9.6315 
  7.7332 
5.9693 
 1.2538
10.
7.
6.
1.25



Respective OF-value is 0.0004196.
The total number of OF-calls is 1375.
N.B !  The  revealed parameters differ from the "true" ones  not because of  RGA-algorithm  shortcomings,
but  because  of  nonuniqueness of the problem.  Again: the value of OF is  0.0004.
As to the RGA-precision, see  the next page




...

To the previous RGA-page. to the previous RGA-page 
 
 
  

  to the next RGA-page  To the next RGA-page
Top
Online Directory of web resources, including Research
Геннадий Рыжиков