A pseudo-incremental rough set-based pseudo outer-product with ensemble learning (ieRSPOP++) fuzzy neural network

Abstract

Attribute and rule reduction through rough set theory in order to increase semantic interpretability is achieved in the rough set pseudo outer-product fuzzy neural network (RSPOP FNN); to improve the capability of RSPOP to function as an incremental system, an incremental ensemble RSPOP FNN is proposed. An incremental ensemble RSPOP FNN, named ieRSPOP, has been previously implemented; however ieRSPOP suffered from some problems: Hebbian Learning, multiple knowledge bases, storage of a huge chunk of data, and inflexible method to invoke rough set reduction.

The problem with Hebbian Learning is the unidirectional growth of synaptic signals, which is not suitable for time-variant data. The Bienenstock-Cooper-Munro (BCM) theory of learning is a theory that provides for Hebbian, anti-Hebbian learning. The Rough Set theory performs feature selection through the reduction of attributes, but also extends the reduction to rules without redundant attributes. To incorporate, rough set theory in an online fashion, a novel ensemble algorithm is used representative how the brain chunks information. Multiple Trace Theory suggests that the hippocampus allows for chunking and plays an important role in recalling memories, since when old memories are cached away they leave traces in in the hippocampus, which are linked to cortical networks. This is contrary to the standard model, which claims that the hippocampus stores memories only temporarily. Brain-inspired theories such as the Ebbinghaus theory on LTM forgetting and displacement theory, which states that forgetting occurs due to new information arriving are compared with the usual fixed decay model of Long-Term Depression.

This paper proposes a neuro-fuzzy architecture, the incremental ensemble rough-set pseudo-outer product (ieRSPOP++) that uses the BCM theory of online learning, rough set theory, and a novel ensemble learning algorithm inspired from mechanisms of the brain.

The performance of ieRSPOP++ is evaluated and compared against existing incremental and non-incremental architectures, through several time series benchmark experiments and prediction of future stock prices. ieRSPOP++ is also used as a regressor for artificial ventilation modeling and volatility prediction for option trading. The results are promising.