Simplification principles are presented in . It aims to eliminates the less useful variables from the rules of a fuzzy inference system.
The procedure needs a current FIS configuration file and a current data file.
It takes the same parameters as the FPA algorithm (used for generating conclusions), and some specific ones: the tolerated loss of performance (as a percentage of the initial performance index), and the homogeneity threshold. This last parameter is used for checking that the widening of the rule input space coverage resulting from the simplification procedure does not lead to abusive simplifications. It sets the dispersion maximum value for the observed outputs of the items attracted by a rule. If the homogeneity threshold value is two low, only very specific rules will be kept, which will make interpolating difficult in the fuzzy inference system. We advise to start the procedure with a value of 0.5.
The KeepLast option is used by the procedure to not remove the last rule whose conclusion is a given class label, case of a crisp output with the classification flag, or a membership function in the case of a fuzzy output.
A validation file can be specified. In that case, it is used at each
step to decide whether to keep not to keep the simplification resulting from the use of the
The procedure output is summarized in the file result.simple.
Each line is a summary of one attempt to simplify the system. The first column gives the name of the FIS configuration file created during the run. The name is formed by appending to the initial FIS configuration file name the structure , being an index number, the indices given in section 2 and finally the rule base characteristics, see section 3.
At the end of the simplification procedure, a window appears that corresponds to the simplest FIS found.
- 1st column: initial Fis filename
- 2nd column: performance index
- 3th column: coverage index
- 4th column: max error
- 5th column: activation threshold used for computing the coverage index
- following columns: the rule base characteristics