Metacognitive learning algorithm for big data analysis

Abstract

The field of data prediction and forecasting techniques is a research area that has found many applications in this modern world. Linear regression models are often inaccurate to model the non-linear relationships that may exist between the variables of study. As such, many techniques in machine learning such as Artificial Neural Networks have been used to learn the data. This project deals with the field of neuro-fuzzy inference systems, where fuzzy logic is embedded into a neural network in order to deal with the uncertainty aspects of the data relationships. In fuzzy inference systems, conditional rules govern the relationship between the input and output variables through fuzzy membership rules. This project will be focusing on the development of a type-2 interval fuzzy inference system (IT2FIS) in Python, where the membership values of the rules themselves are fuzzy. This is in contrast to type 1 fuzzy inference systems where the membership function is exact.In particular, this project will be layering metacognition on top of the developed IT2FIS. The use of metacognition, which is essentially regulated learning controlling the how-to-learn, when-to-learn and what-to-learn dimensions, will provide significant improvements to the system. This paper will therefore investigate the effects of metacognition layered into the standard IT2FIS that is developed with gradient descent, as well as comparing the systems on benchmark problems that is widely used in literature. Overall, metacognition has been found to be a very effective method in reducing the computational time and reducing the error significantly on many types of benchmark problems. However, more research and effort could also be done to apply metacognition to inference systems that use superior learning mechanisms to the gradient descent backpropagation mechanism.