Optimization

Java Interface: Learning menu, Optimization submenu: Two options:

  1. Custom Solis & Wets option.
  2. Standard Solis & Wets option.

Command line, loopoptims program:

The loopoptims program requires the following arguments:

default=optimize rule conclusions

Optional parameters:

-b the basename of the resulting FIS configuration file;(default optim)
-it the number of iterations; (default 100)
-e the range of the gaussian noise; (default 0.005)
-v the max number of constraints violations before it counts as an iteration (default 1000)
-f the max number of failure steps in an algorithm (default 1000)
-d the equality center distance threshold (default 1e-6)
-ns Number of sample pairs to generate from data file (default 10-put -ns 0 for no sampling)
-cs Draw samples with respect to class ratio in data file (class is last column, default no))
-rs Ratio learning/all pairs (default 0.75, maximum 0.9)
-s integer to set seed value for parker miller random;(default 0:new sampling each time);
-in "x y ", the string of input numbers to optimize (starting at 1, order is important, default no input optimization)
-r optimize rule conclusions
-o optimize fuzzy output (default false)
-n Output number to consider (default 0: first output);
-m minimum membership (default 0.1)
-l1 Solis Wetts Constant 1 (default 0.4)
-l2 Solis Wetts Constant 2 (default 0.2)
-l3 Solis Wetts Constant 3 (default 0.5)
-u Number of loops for optimizing (default 2)
-c relative tolerated loss of coverage (default 0.10 ; 1.=10.0%)
-nc minimum distance between MF centers (default=1e-3)
-g create intermediate files (default false)
-a for intermediate display (default false)
-wl for wordless.(default is not wordless)
-nc distmin to impose a minimum distance between the neighboring MF kernels (default 0.001).

Command line example:

loopoptims rice.fis rice -ns 2 -s 102 -b optimfis

which optimizes the rule conclusions for the FIS configuration file rice.fis, with a cross-validation procedure including 2 pairs of learning(75%)-test (25 %) samples, and stores the optimized FIS for each learning sample, in the optimfis-lrn.sample0-final.fis and optimfis-lrn.sample1-final.fis files. The med.fis file is also built using the median parameters of the optimized FIS.

or

loopoptims rice.fis rice -ns 2 -s 102 -b optimfis -in '1 3' -r

which optimizes the fuzzy set parameters of the input variables 1 and 3, as well as the rule conclusions for the FIS configuration file rice.fis. The generated files have the same names as above.

A perf.res file is written, that includes the accuracy of the initial FIS, of each optimized FIS and of the medain FIS, calculated for the initial dataset, each test sample and each learning sample.