Hi,
I have developed a fuzzy control system that produces an overall accuracy of 90.5% and I want to be able to prove that the fuzzy rule base was not tuned specifically for the data set in order to fabricate the accuracy.The rule base was developed purely through intuition.
The data set is a publicly available annotated corpus that is used by machine learning researches. I used the annotators label and the output of my system in order to figure out the accuracy, that is to say if majority of the annotators agree on a certain label for an input and the output of my FCS matches this label then I consider it as a correct outcome.
Is there any way to prove that my fuzzy rule base is in no way fabricated to match this particular dataset? I'm asking since the rule base was developed using intuition. and is there a standard way to test my system given a dataset like above.
Hi Keshav Kumaresan,
In this context, to fabricate is to invent the "fuzzy rule base" in order to deceive. What is there to be deceived in the first place?
Fuzzy Logic systems are universal approximators. If the data matches certain conditions, the fuzzy system can approximate the data with high accuracy, whether it is built intuitively, or tuned by ANFIS.
However, I have encountered cases where some researchers claimed that a novel (triangular MFs-based) fuzzy system is built when it performed exactly the same like "f(x) = x", a linear function.
Depending on the applications, in some cases, these "linear" fuzzy systems are deemed without merits because it would be much easier to use "f(x) = x" directly.
For example, consider the following linear system.
x' = A*x + u
Naturally, a proportional (P) controller is sufficient to stabilize it.
u = - K*x
One type of the fuzzy P control would like
u = - K*𝓕(x)
where 𝓕(x) is the triangular MFs-based fuzzy system that produces exactly the "x" function.
To create fuzzy controller or fuzzy system, you have acquired a fuzzy rule table experimentally. I suggest read more in (Monotonicity of fuzzy rule bases) this help to evaluate the relation among fuzzy rule and it help to have a evidence about how your FLC is Convergent the truth intuitively.
Hi Keshav Kumaresan,
I'm figuring out what you are trying to tell. Some researchers have developed a machine learning controller.
You have successfully incorporated the machine learning expert knowledge into the fuzzy system and achieved the performance accuracy of 90.5%
Are you implying that you fear the researchers would accuse you of "fabricating" the fuzzy system?
Hi Yew-Chung Chak ,
Thank you for your answer, and yes that is exactly what I'm asking, I would like to know if there is a standard/ mathematical way that I can prove my system is not in any way fabricated. That is if someone raises the question that my rule base has been tuned only for that particular dataset and hence the high accuracy and it won't work given another dataset. Is there any standard way to refute that claim?
Thank you.
Hi Keshav Kumaresan
In that case, you can pick a few different initial conditions and show that your fuzzy controller works in different scenarios.
- Badges
- Science topic