Automatic modularization with speciated neural network ensemble

Decomposing a complex computational problem into sub-problems, which are computationally simpler to solve individually and which can be combined to produce a complete solution, can efficiently lead to compact and general solutions. Neural network ensemble is one such modular system that uses this divide-and-conquer strategy. Diverse set of networks improves ensemble's performance over its constituent networks. Artificial speciation is used here to produce this diverse set of networks that solve different parts of a data classification task and complement each other in solving the complete problem. Fitness sharing is used in evolving the group of neural networks to achieve the required speciation. Sharing is performed at phenotypic level using modified Kullback-Leibler entropy as the distance measure. The group as a unit solves the classification problem and outputs of all the networks are used in finding the final output. For the combination of neural network outputs 3 different methods - Voting, averaging and recursive least square are used. The evolved system is tested on two data classification problems (Heart Disease Dataset and Breast Cancer Dataset) taken from UCI machine learning benchmark repository.