The evolving mamdani fuzzy inference system (eMFIS) with its applications in straddle option trading and ovarian cancer classification

• Main Author: Mario Hartanto
• Other Authors: Quek Hiok Chai
• Format: Final Year Project
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies
• Online Access: http://hdl.handle.net/10356/61563
• Institution: Nanyang Technological University

Neuro-fuzzy systems (NFS) are hybrid systems which benefit from the expressive IF-THEN fuzzy rules and the learning ability of neural network. As evident in recent researches, there are growing interests in NFSs with online learning capabilities, as opposed to the traditional NFSs which learn through offline batch learning mechanism. However, many online learning systems, when used to model data in time-varying environments, are not able to automatically detect and react to the occurrence of concept drifts and shifts, raising questions to their life-long learning capabilities. While there have been some proposed approaches to deal with concept drifts, such as through ensemble learning and moving window, these approaches may suffer from: (1) interpretability issue from multiple knowledge bases, (2) difficulty in choosing appropriate parameter for the window, (3) concept drifts/shifts detection may be heuristic or susceptible to false positive. This paper proposes a novel neuro-fuzzy system architecture called evolving Mamdani Fuzzy Inference System (eMFIS), which can overcome the above issues, in order to realise an online system with a semantically expressive fuzzy rule base, which is capable of life-long learning, even in time-varying environments with the presence of concept drifts and shifts. eMFIS applies the Bienenstock-Cooper-Munro (BCM) learning theory to self-reorganise its rule base and keep it updated with the most recent knowledge, which is representative of the environment. eMFIS uses a concept of cluster neighbourhood age to automatically detect the occurrence of drifts and shifts. eMFIS also incorporates a novel proposed fuzzy clustering algorithm, called 2-Staged Incremental Clustering (2-SIC), which is inspired from the development of human categorical learning mechanism as human grow from infancy to adulthood. 2-SIC enables an incremental fuzzy clustering to be done without any prior knowledge of the data and without parameter determining the widths of the clusters. eMFIS is benchmark against existing architectures in both classification and regression problems, using both time-invariant and time-variant data. eMFIS shows particularly strong forecasting performances when modelling time-series data and real-life stock index prices. eMFIS is further used for volatility prediction for straddle option trading and ovarian cancer classification. The results are encouraging.