Machine Learning in Real-Time with Fuzzification

Files here are working and tested code and are used in the MaRz project. Some aspects of the project are still in development, but everything here is working and not testing files.

• numpy for dataset management
• unittest for testing
• sklearn for splitting and analysis

• matplotlib.pyplot and seaborn for charts
• tensorflow for heart disease comparison

The process of making a query with MaRz happens in three phases: Preprocessing, Hyper-boxing, and Querying. For full tests, entire small (~5000 lines or fewer) datasets can be processed all together.

In order to ensure compatibility, datasets should be formatted as 2D numpy arrays with rows of inputs and columns of attributes. The last column of the dataset should be the target outputs of the data.

The preprocessing done in dataset_preprocessing.generate_index_table creates a sorter table of indices of the dataset and fills it with columns where-in each column is the indices of the corresponding column of the data set, in the order they would be in if the dataset were sorted (stable) by that column.

The second part of preprocessing is to generate the base fuzzy width for the dataset. This is done with dataset_preprocessing.get_base_fuzzy, which takes a properly formatted dataset and returns a list of value ranges for each feature of the dataset, referred to as the base_fuzzy.

At this point, an input is needed. An input is a list of features the same size as a single entry in the dataset, but without a target on the end.

Generating an appropriate hyperbox for the data is a small tuning process, which uses the get_alpha function in get_alpha_sorted.py. This version uses an index table as generated in the preprocessing. The unsorted version does not use an index table, but it less efficient.

The alpha value is a number between 0 and 1 by which each number of the base_fuzzy is multiplied in order to create a hyperbox around the input within the dataset which contains the input as well as a minimum of n additional datapoints, as indicated by the points argument. get_alpha also returns a container of the indices of the points within the hyper-box.

To query MaRz with the chosen input uses the marz_get_output.get_output function, which applies the fuzzy calculation to the points in the hyper-box and produces a prediction for the input. This output is a decimal value appropriate to the targets of the dataset.

The run_dataset.py file makes it convenient to process a full dataset and get back two lists containing the real and predicted values returned when each line of a dataset is given as input to the MaRz process. These output lists can be used to calculate an accuracy score, with the entire dataset as "testing split." See this in code in the airfoil_data.py experiement.

In order to handle columns where every value is the same, the base_fuzzy for that column is converted from 0 to 0.000001, to prevent division by 0 downstream.

This code is currently under anonymous submission for the 2024 AAAI Conference. External sources are cited where they are used.