In order for the generalized bell membership function to retain its defined shape and domain, two restrictions must be placed on the b parameter: 1) b must be positive and 2) b must be an integer. Using backpropagation to tune the membership parameters (a,b and c in the bell), it appears to be possible that the correction to b will break one or both of these restrictions on b. Can someone please explain to me how we can use backpropagation to tune the b parameter (as well as a and c) without violating 1) and 2)?
After re-reading Jang's original (1993?) paper on ANFIS, I learned he recommended simply squaring the b-parameter to deal with the notion that b could be changed to a negative value when using back-propagation. While this solves the negative domain issue, the issue still remains that if b is a non-integer, then the bell function loses its intended shape. I suppose after squaring one could round to the nearest integer, but I suspect that doing so might hinder convergence. Perhaps a consideration is only tuning the a and c parameters using something like a genetic algorithm.