Strategies for Evolutionary Data Driven Modeling in Chemical and Metallurgical Systems

Abstract

Engineering problems often deal with industrial, experimental, or simulated data containing a very significant amount of random noise. To construct the right model out of such information is often a very cumbersome task and either over or underfitting of the data is a common occurrence. Complex interaction between the variables adds to further problems. With the advent of multi-objective genetic and evolutionary algorithms, some modeling strategies have been proposed in recent times that are capable of constructing the right tradeoffs between the accuracy and complexity of such models. Also, those strategies are able to establish the nature of interaction between the variables in an intuitive but effective way, without involving unnecessary mathematical complications. Evolutionary neural network (EvoNN) and BioGP (bi-objective genetic programming) are two recent algorithms developed to accomplish this, by altering several features of the conventional neural networks (ANN) and genetic programming (GP) through a predator–prey type bi-objective genetic algorithm (GA), which gives rise to a Pareto frontier between the accuracy and complexity of the models, and a decision maker (DM) is allowed to select an appropriate model even by applying some additional criteria, if necessary. Those are discussed in detail along with some pruning algorithms that are found to be highly effective for constructing data-driven models and their subsequent optimization. Few commercial software are also discussed along with several real-life application in the fields of chemical and metallurgical engineering.