Big Bang Big Crunch Based Development of Machine Learning Models: A Soft Computing Based Novel Approach

Abstract

This paper proposes a new metaheuristic approach to design and develop Artificial Neural Network (ANN) based machine learning models. The proposed approach is a data-driven technique that works in three phases. Phase 1 of the proposed algorithm evolves the optimal architecture of an ANN model. In this phase, we evolve the number of hidden layers and the number of neurons in every hidden layer of the ANN model. Phase 2 is the model optimization phase in which we optimize the weights of the ANN architecture, evolved in phase 1. To evolve the architecture and weights for the specific ANN architecture, we used Big Bang Big Crunch (BB-BC) optimization algorithm. In Phase 3 of the approach, the optimized model is implemented. For implementation purposes, the proposed approach writes the code for the optimally evolved ANN model. We validated our proposed approach by developing machine learning model from a nickel cadmium battery charging training data set and compared its performance with manually configured ANN, Multi-Regression, and KNN (K Nearest Neighbors) based machine learning models. We observe that the proposed BB-BC based automated approach completely outperforms the other three machine learning models.