Generic Hebbian ordering-based fuzzy rule base reduced neuro-fuzzy system with fuzzy rule interpolation (RS-Hebb+)
Date of Issue2017
School of Computer Science and Engineering
Centre for Computational Intelligence
Neuro-fuzzy system, traditionally used in dynamic data sets modelling, is now evolving rapidly in both structure and style, including the trends from offline system changing to online system, and increasingly more concepts added to address issues like data sparsity, time-variants and non-linearity. As theoretical researches go on, the results have also been applied in a wild range of industries, including traffic control, rainfall prediction and financial markets. Nevertheless, most of the systems now existing are not capable enough to dealing with sparse and dynamic time series data, which can be widely found in stock markets. Besides the other popular techniques introduced to address this issue, interpolation and extrapolation are the ones that gives most promising results and potential improvements on the existing system. Inspired by the paper “Fuzzy interpolation and extrapolation: a practical approach” (Huang and Shen, 2008) , this paper is built on the base of Liu’s work in  (RS-Hebb) with interpolation/extrapolation techniques to address the data sparsity. This improvement enable the original system by Liu to deal with sparse data as concept drift and shift happens. The RS-Hebb mostly address the trade-off of two requirements of neuro-fuzzy modeling: firstly, interpretability, which means to what degree the behavior of the system can be described in an interpretable way; Secondly, accuracy, that whether the outcome can be approximated accurately. The proposed model is named as Hebbian Ordering-Based Fuzzy Rule Base Reduced Mamdani Neuro-Fuzzy System with Fuzzy Rule Interpolation (RS-Hebb+). The Hebb rule reduction balance interpretability and accuracy by interactively parameter tuning. Hebbian ordering, representing the percentage of samples covered by each rule, is used when merging membership functions(MF) and number of rules. Only MFs that above certain threshold of similarity measure is kept, when others are merged and deleted from rule base. Inherited the properties of RS-Hebb, RS-Hebb+ has the following advantages: (1) The rule base generated is intuitive, consistent and compact, which includes a reduced attribute set, well-separated membership functions in every dimension of inputs, and comparably small number of rules. (2) Trade-off between interpretability and accuracy is well balanced, by Hebb rule reduction. (3) It is capable for dealing with concept drift and shift through interpolation when the data is sparse. The proposed RS-Hebb+ model is benchmarked against several existing systems by applying to datasets with different properties. The model is then applied in a well-established trading system for stock market, showing promising and encouraging results.
DRNTU::Engineering::Computer science and engineering::Mathematics of computing::Probability and statistics
Final Year Project (FYP)
Nanyang Technological University