ANFIS Model, tips about improving performance.

I have a question regarding about improving the performance of an ANFIS (adaptive neuro Fuzzy inference system) model.

In MATLAB, I have been training a model with 5 inputs, with 816 data point for each input, and 1 output, with 816 data points. The best model has the following properties:

• 2 memberships function for each input, pi-shaped.
• The output has constant membership function for each of the 64 rules.

The training parameters are the following:

-Epoch: 500

- Initial step size: 0.1

- Step size decrease rate: 0.9

- Step size increase rate: 1.1

- Optimization method: Backpropagation with gradient descent.

- Clustering method: grid partition

This is how the model output behaves:

From 0 to 568 Dias, is how the model behaves (red line) with the training dataset (blue), while from 569 to 815 Dias is the validation set.

Any kind of help is aprecciated.

Answer 1

[Sam Chak]

Hi @Jocelyn Matus Ancavil

ANFIS typically performs optimally with continuous datasets. However, your dataset seems to consist of discrete values {0.0, 0.3, 0.5, 0.7, 1.0}. Consequently, it requires post-processing to enhance the predictive quality of the ANFIS model. A straightforward approach could involve rounding off the prediction values based on user-defined rules, such as 'rounding to the nearest.'

It seems you've allocated the first 70% of the data (first 568 days) for training and the remainder for cross-validation. However, I would suggest opting for a cross-validation approach where you select one day every three days. This should provide ample spread across the days for training purposes.

Certain day intervals exhibit rapid changes. Using only two membership functions per input may not adequately capture the dynamics of these rapid changes. Increasing the number of membership functions to 3 or 4 could potentially enhance prediction accuracies.

[Jocelyn Matus Ancavil]

Hi! Thanks for the answer, it helped me a lot. I want to ask if you have any book or paper recommendations about ANFIS or fuzzy logic optimization, or about improving these kind of models.

[Sam Chak]

Hi @Jocelyn Matus Ancavil,

Thank you for your support.

I believe there is a single book that delves into ANFIS in great detail, titled "Neuro-Fuzzy and Soft Computing." The first author of the book is considered the pioneer of ANFIS, as he developed the original fuzzy modeling tool based on neural networks in the early 1990s. Presently, in MATLAB, you can also fine-tune and optimize fuzzy systems using built-in algorithms like the genetic algorithm, particle swarm optimization, pattern search, and simulated annealing.

Around six months ago, I provided assistance to an individual in this thread to train ANFIS for discrete data. Please take a moment to review it. If you find any valuable insights in my three Answers, I would appreciate it if you could consider voting them with a thumbs-up (👍) as a token of support.