• Mathematical background: Jang, Sun, and Mizutani, "Neuro-Fuzzy and Soft Computing".

• Datasets: UCI Machine Learning Repository.

• The code has been written and tested in Python 3.7.7.
• Multi-input/multi-output (multivariate) adaptive neuro-fuzzy inference system (ANFIS) implementation for regression and classification.
• Quadratic cost function for continuous problems and cross-entropy cost function for classification problems.
• Classes in classification problems are determined automatically.
• Sigmoid and cross-entropy function are computed using a numerically stable implementation.
• Generalized Bell curves depending on three parameters (mean, standard deviation, and exponent) are used as premise membership functions.
• Hyperplanes depending on the number of features are used as consequent functions.
• A particle swarm optimizer (PSO) is used to solve the minimization problem. More info about it here.
• Limits/constraints on the parameter values (similar to regularization in neural networks) can be easily done through the PSO boundary arrays.
• The ANFIS class in ANFIS.py is not constrained to the PSO solver but it can be used with any other optimizer not gradient-based.
• File utils.py consists of several utility functions, including an helper function to build the PSO boundary arrays.
• Usage: python test.py example.

example Name of the example to run (plant, stock, wine, pulsar.)

problem Defines the type of problem. Equal to C specifies a classification problem, anything else specifies a continuous problem. The default value is None.

split_factor Split value between training and test data.

data_file File name with the dataset (csv format.)

n_mf List, tuple, or array with the number of premise functions of each feature.

nPop, epochs Number of agents (population) and number of iterations (PSO.)

mu_delta Allowed variation (plus/minus) of the mean in the premise functions. It is given as fraction of the corresponding feature data range.

s_par Center value and allowed variation (plus/minus) of the standard deviation in the premise functions. The center value is scaled based on the corresponding feature data range.

c_par Range of allowed values of the exponent in the premise functions.

A_par Range of allowed values of the coefficients in the consequent functions.

For the meaning of the PSO parameters K, phi, vel_fact, conf_type, IntVar, normalize, and rad, see here and file pso.py.

There are four examples in test.py: plant, stock, wine, pulsar. The values common to (most of) all are:

problem = None
split_factor = 0.70
K = 3
phi = 2.05
vel_fact = 0.5
conf_type = 'RB'
IntVar = None
normalize = False
rad = 0.1
mu_delta = 0.2
s_par = [0.5, 0.2]
c_par = [1.0, 3.0]
A_par = [-10.0, 10.0]

data_file = 'plant_dataset.csv'
n_mf = [1, 1, 1, 1]
nPop = 40
epochs = 500

Original dataset: https://archive.ics.uci.edu/ml/datasets/Combined+Cycle+Power+Plant.

The dataset has 4 features (inputs), 1 label (output), and 9568 samples.

The ANFIS has a layout of [1, 1, 1, 1] and 17 variables.

Predicted/actual correlation values: 0.965 (training), 0.961 (test).

data_file = 'stock_dataset.csv'
n_mf = [2, 2, 2]
nPop = 100
epochs = 500
A_par = [-1.0, 1.0]         # Supersedes the default A_par

Original dataset: https://archive.ics.uci.edu/ml/datasets/ISTANBUL+STOCK+EXCHANGE.

The dataset has 3 features (inputs), 2 labels (outputs), and 536 samples.

The ANFIS has a layout of [2, 2, 2] and 82 variables.

Predicted/actual correlation values: 0.883 (training), 0.871 (test).

data_file = 'wine_dataset.csv'
problem = 'C'
n_mf = [3, 2]
nPop = 40
epochs = 500

Original dataset: https://archive.ics.uci.edu/ml/datasets/Wine+Quality.

The dataset has 2 features (inputs), 6 classes (outputs), and 1599 samples.

The ANFIS has a layout of [3, 2] and 123 variables.

Predicted/actual accuracy values: 58.2% (training), 59.8% (test).

data_file = 'pulsar_dataset.csv'
problem = 'C'
n_mf = [3, 4, 2]
nPop = 40
epochs = 200

Original dataset: https://archive.ics.uci.edu/ml/datasets/HTRU2.

The dataset has 3 features (inputs), 2 classes (outputs), and 17898 samples.

The ANFIS has a layout of [3, 4, 2] and 219 variables.

Predicted/actual accuracy values: 97.9% (training), 97.7% (test).

The initial and final premise membership functions for this example can be seen here.