Well, this is very difficult to know without knowing the application and seeing the results and how the controllers were tuned. But basically, yes. This is my experience: I could always find parameters for a PID such that it performed better than the FLC. Of course, in the applications, where both controllers can be used.

The PID can perform well on systems where the uncertainty is bounded. On the other hand, the FLC will be useful in non-linear systems with a lot of uncertainty.

The performance of any controller depends on , what desirable objective one want to fulfill, i.e, tracking, stabilization, optimization etc. What fuzzy does better is the task of uncertainty handling......it may perform less in ideal environment but perform satisfactory when disturbances are present (already proved in many articles). Two important points are:

1- If one is comparing PID and FLC, one should compare the performance for various cases (ideal, uncertainties, noise) etc.

2- FLC have many tuning parameters and as it is based on ruled based modelling, it is important to try different membership functions, different rule set, different fuzzifier/defuzzifier etc.

Hi Hany,

It seemed that you faced this problem six months ago. Higher IAE value may indicate large overshoot or long settling time. If a process system can be satisfactorily regulated by a PID controller, then it is 100% possible to design a simple fuzzy controller that performs as good as the PID controller does. You may consider this kind of fuzzy controller works like a PID control emulator. It works because fuzzy logic systems are “universal approximators”, something like cutting a circular cake into 6 almost equal slices without using a drawing compass.

However, if you design a fuzzy controller intuitively from scratch (based on your observation or experience) without relying on previous PID control design work for reference, then the fuzzy controller may underperform at first. In such a case, it is not uncommon for a well-tuned PID controller to outperform a heuristically-designed fuzzy controller, since the PID gains can be tuned optimally using various control design tools.

Hence, in order to build a fuzzy controller that perform at least on a par with the PID controller, you should first design and tune a PID controller, and then design a simple fuzzy controller that serves as the PID controller counterpart using the knowledge from the tuned gains (Kp, Ki, Kd). If successful, then you can find out whether a fuzzy controller will perform better by gradually adjusting the membership functions and modifying the fuzzy rules. Some control theorists use evolutionary algorithms to optimize the performance of the fuzzy controller.

The Fuzzy PID control design philosophy is well outlined in the works of Guanrong Chen & Trung Tat Pham, Zdenko Kovacic & Stjepan Bogdan, and Jan Jantzen. Imagine that the initial design of a fuzzy controller cannot perform as good as the PID controller in the first place, how do you find hope that a tuned fuzzy controller will outperform the PID controller?

Would you like to talk about the initial design of your fuzzy controller (by heuristic or systematic)?

P.S. If you find it too late to grasp the design philosophy at this point of research, there is a workaround! Since you have the input-output data of the PID-controlled system, then you probably can use the built-in ANFIS tool (adaptive neuro-fuzzy inference system) in MATLAB to create a fuzzy logic controller. This might be an alternative solution which complies with the PID performance requirements. The ANFIS-designed fuzzy controller probably just needs a little tweaking on the membership functions and IF-THEN rules to yield a lower IAE than the PID controller.

Dear Dr Yew-Chung Chak,

Thank you for your kind answer. Yes, actually, I have faced this problem a few months ago. Actually, I used knowledge from PID tuned gain to build my fuzzy controller, and here I got the simulation result. I still working on my Fuzzy, since I think it might be under-performed.

But now, I think maybe I should compare both PID and FLC for various cases (ideal, uncertainty, etc), as been suggested by Dr Madan Mohan Rayguru. And also I should try ANFIS. Thank you very much for the tips. I really appreciate it..

Hi Hani,

Sorry for the late reply. Keep in mind that the initial fuzzy controller should to be designed with the intention to emulate the performance of the PID controller (and not to perform better than PID controller on the first attempt). In order to design a fuzzy controller that works like a linear PID controller, the Sugeno-type fuzzy inference and piecewise linear membership functions must be used.

Okay, here is a simple demo to show you that it is possible to build a fuzzy controller that performs as good as the linear PID controller does. The fuzzy controller in this context can be created using the ANFIS tool. Hence, no theoretical explanation needed.

Run the simulation (see attached Simulink slx file) and you'll find the performance of the FLC (IAE = 1.034) is practically indistinguishable from the PID (IAE = 1.03). Note that each fuzzy input consists of a basic set of 3 triangular MFs. If higher accuracy is required, then more number of MFs should be used.

First, you have to run your PID control system so that it generates and sends the PID input and output data (performance which you want the FLC to emulate) to the Workspace. The steps for training a Neuro-Fuzzy system is given as follows:

• Enter this in the command window: "data = [in out];"
• Run the Neuro-Fuzzy App (ANFIS tool).
• Load Data from the Workspace.
• Generate FIS using triMF for INPUT and linear MF for OUTPUT.
• Set Epochs to 30 and start the training. Increase the Epochs if the training result does not converge.
• Test ANFIS-trained FLC which compares against the PID data.
• If satisfied, press Ctrl+1 to bring up FIS Editor, or else Train again with another setting.
• Export ANFIS-trained FLC to File and then to Workspace.

Now you have a reliable FLC. Replace the existing PID controller with the ANFIS-trained FLC and then compare its performance.

One important thing you have to know is that using ANFIS tool to create a FLC does not automatically make your work novel. One might draw an analogy between the ANFIS tool and a CASIO calculator that is used to perform tedious arithmetic operations. Furthermore, the ANFIS-trained FLC only performs as good as the PID controller does.

Good luck!

For FLC based control system, you may need to optimize the MFs in order to achieve a better performance like this paper posted on https://goo.gl/HrFxnY

Hi Zahid,

Thanks for your comments and observations. As I mentioned in my first post, the ANFIS approach in the demo is just a workaround for Hani to construct a fuzzy controller that performs as good the designed PID controller. Because the input/output data is based on a fixed reference step input, therefore, the ANFIS doesn't work for setpoints other than 1.

Well, to build (not design) a fuzzy controller for a range of setpoints, say 1 to 5, you need to train a set of data that covers the operation range. In the attached simulink file, I used a set of unicycle slow-varying sinusoidal input/output data (0 to 6) for ANFIS training. Hence, a new set of fuzzy rules is obtained. In the multi-graph, the ANFIS-based fuzzy controller actually performs better than the PID controller. You can verify that through the IAE values.

It is normal if the system you are controlling does not require adaptive or fuzzy control. What tunning method you are using in FLC ?

Fuzzy logic systems are universal approximators. Therefore, it is certainly abnormal and uncommon for a model-based Fuzzy controller to be underperforming when it is designed and built correctly. In other words, as long as the control law can be expressed mathematically, then the control law can be 'fuzzified' to produce the result to a good approximation.

Hi, You can check for comparison my research on an application of FLC controller in robotics: "Research of an Omnidirectional Mecanum-Wheeled Platform with a Fuzzy Logic Controller".

Hi Łukasz Rykała, thanks for sharing your paper on the fuzzy control design. I just want to add that the indentation on the control surface, generated by your Mamdani FIS, can be reduced through the fine shaping of the corresponding MFs in the antecedent of the FIS. Also, the gains ke, ki, and ku on pp. 3, 6 are not found in any equation in the paper. Could you enlighten us?

Thanks for the tip.

Mentioned gains are used as a scaling factors for the input signals (ke, ki) and output signal (ku) of the FLC. Their values have been determined experimentally after carrying out many simulations. Exemplary values: ke=1/80, ki=1, ku=50.