I am stuck on a classification problem using fuzzy inference system. I have read that fuzzy logic can be used for classification problems. But how to construct the output membership function for non interpolating classes. For example. an image of a tiger or a bike. If these two have no interpolating properties, then how do I construct the output membership function? Please help me in solving my problem. Thanks all in advance.
Genetic algorithms might be used to tune the membership function parameters. Here, one would start with a population of random set of parameters for membership functions, say 100 complete sets. Then test each set with your classification data set. Use a norm measure for the output of each parameter set and your desired correct output, then discard all but (for example) 10% of the parameter sets with the lowest error. Replace the remaining 90% of parameter sets with slightly stochastically perturbed versions of the 10% best performing set. Repeat this process until you have a satisfactorily performing membership function set. Of course, genetic algorithms of this nature need to be implemented via computer using some sort of batch processing.
@andrew ....Thanks for your reply. My doubt is not regarding tuning. Its about initial stage when we construct the membership function. This is the basic doubt which I want to get clarified. If its a classification problem then it has to belong to either a class 0 or class 1 or class 2 or so on. In that case if I take a triangular membership function with each peak at either 0 or 1 or 2, then does it violate any rule. A figure has been attached based on this assumption. Here, there are 3 classes, tiger, cycle and tree. Here can I say that if the crisp output value comes to be within 0-0.5 then its class 0, 0.5-1.5 then its class 1 and 1.5-2.5 then its class 2. I formulate the rule base accordingly.
Use singleton output membership functions corresponding to the available classes and either
- change the inference technique to a single rule based one (using only that rule whose firing strength is the biggest)
or
- use a defuzzification method that ensures that the output class will be the one with the maximum membership value
Thanks Robert, Zsolt and Saravanan. Where can I find a proper tutorial on the use of fuzzy singleton for classification purpose? This is because I am not able to visualize how my problem can b solved.
- Badges
- Science topic