GEMM-eMFIS(FRI/E) : a general episodic memory mechanism for neuro-fuzzy networks

• Main Author: Tekwani Puneet
• Other Authors: Quek Hiok Chai
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/148054
• Institution: Nanyang Technological University

Machine learning models can be used in fields like finance, engineering, medicine and science to make real time predictions, but this can become a bit challenging whenever there is a sudden and large change in the data from a steady state, known as transient behavior. An example of such a behavior would be when a stock’s price suddenly shifts [1] upon corporate announcements, global recessions or any other events in the market which would be of interest to everyone. The idea is to build a computationally intelligent system (CIS) to keep a memory of these shocks or transient events, learn the most important rules from these events and reuse these rules when any similar event takes place. This concept in humans stems from the theory of episodic memory [2] which allows the storage and recall of similar events, in order to take appropriate actions. Neuro-Fuzzy Systems (NFS) integrate Fuzzy Inference Systems (FIS) and Artificial Neural Networks (ANN) to provide the best learning complex representation of data and also maintaining human interpretable sets of rules [3]. They are known to perform well in financial domain [4, 5] because of their ability to learn rules in real-time and self-evolving structures which make them suitable for realizing an episodic memory mechanism. However, existing NFS possess the ability to deal with real-time updates in data but they struggle during periods when there is a sudden wild change in data i.e., the price fluctuates wildly (transient behavior), such events could be recessions, news or corporate actions. To deal with transient behavior in the stock market is challenging and not many models seem to possess this capability. This paper proposes a novel general online episodic memory mechanism which demonstrates an instance of its integration into Neuro-Fuzzy architecture called Episodic Memory Based, evolving Mamdani Fuzzy Inference System with Fuzzy Rule Interpolation and Extrapolation (GEMM-eMFIS (FRI/E)). It is inspired by established theories that govern human memory such as Episodic Memory [2], Multi-Store Model [6] and Autobiographical Memory [7]. It learns from past events, by storing and retrieving them from an episodic memory cache during event-driven transient behavior, boosting performance, while potentially reducing the number of rules needed. GEMM-eMFIS (FRI/E) [4, 5] also has several in-built mechanisms that enable to learn effectively from continuous stream of online data namely (1) Bienenstock NTU School of Computer Science & Engineering 2 GEMM-eMFIS (FRI/E) Tekwani Puneet Cooper Munro (BCM) learning theory to keep its rule base updated; (2) 2-Staged Incremental Clustering; (2-SIC) algorithm to determine cluster width; (3) Interpolation and Extrapolation of rules with to deal with concept shifts and drifts in the time-variant data; (4) Rule pruning and merging to keep the rule base compact. GEMM-eMFIS (FRI/E) is benchmarked against other NFS on various time-variant datasets such as stock index prices, and shows encouraging results and strong forecasting performances.