P. Arabshahi et al., POINTER ADAPTATION AND PRUNING OF MIN-MAX FUZZY INFERENCE AND ESTIMATION, IEEE transactions on circuits and systems. 2, Analog and digital signal processing, 44(9), 1997, pp. 696-709
A new technique for adaptation of fuzzy membership functions in a fuzz
y inference system is proposed, The painter technique relies upon the
isolation of the specific membership functions that contributed to the
final decision, followed by the updating of these functions' paramete
rs using steepest descent, The error measure used is thus backpropagat
ed from output to input, through the min and max operators used during
the inference stage, This occurs because the operations of min and ma
x are continuous differentiable functions and, therefore, can be place
d in a chain of partial derivatives for steepest descent backpropagati
on adaptation, Interestingly, the partials of min and max act as ''poi
nters'' with the result that only the function that gave rise to the m
in or max is adapted; the others are not, To illustrate, let alpha = m
ax [beta(1), beta(2), ..., beta(N)]. Then partial derivative alpha/par
tial derivative beta(n) = 1 when beta(n) is the maximum and is otherwi
se zero, We apply this property to the fine tuning of membership funct
ions of fuzzy min-max decision processes and illustrate with an estima
tion example, The adaptation process can reveal the need for reducing
the number of membership functions, Under the assumption that the infe
rence surface is in some sense smooth, the process of adaptation can r
eveal overdetermination of the fuzzy system in two ways, First, if two
membership functions come sufficiently close to each other, they can
be fused into a single membership function, Second, if a membership fu
nction becomes too narrow, it can be deleted, In both cases, the numbe
r of fuzzy IF-THEN rules is reduced, In certain cases, the overall per
formance of the fuzzy system ran be improved by this adaptive pruning.