Multi-Model Consensus Supported by Logical Inference System

Abstract

Automatic classification is a vital research topic and was intensively studied. From the mathematicians’ point of view, model accuracy enhancement is the ultimate result. Many researchers reported high accuracy of their models after modifying or duplicating some training data samples to improve accuracy. Data modification before building the mathematical model means that the model is data sensitive. In this paper, different learning models are trained, without modifying the dataset properties, while their decisions were considered as input to the logical consensus paradigm. Every different learning criterion classify the patterns with different error types. Therefore, using their final results as training inputs to another, rather simple layer, increased accuracy dramatically. The suggested approach abstracts the decision made by different learning models, passing those models behavior to a binary logical decision layer, which produced the results. The training samples, the decision abstraction layer and the final decision were used to update the overall logic of the proposed system. During the training phase, the update logical decision module will keep updating its parameters. After all, the proposed paradigm of combining different learning approaches with a simple binary circuitry achieved almost 100% accuracy.