Data-driven paradigms of EvoNN and BioGP

This paper will present the operational details of two recent algorithms EvoNN (Evolutionary Neural net) and BioGP (Bi-objective Genetic Programming) which are developed for modeling and optimization tasks pertinent to noisy data. EvoNN uses a neural net architecture while BioGP is based upon a tree structure typical of Genetic Programming. A bi-objective Genetic Algorithm acts on a population of either trees or neural nets, seeking a tradeoff between the accuracy and complexity of the candidate models, ultimately leading to the optimum models along a Pareto frontier. Both the paradigms are tailor-made for constructing models of right complexity, and in the process of evolution they exclude the non-essential inputs. By default, an optimum model satisfying the Corrected Akaike Information Crtierion (AICc) is recommended in case of EvoNN, and for BioGP the optimum model with the minimum training error is recommended. However, a Decision Maker (DM) can select a suitable model from the Pareto frontier by appropriate one can be easily picked up by applying some external criteria, if necessary. Both the algorithms tend to avoid overfitting or underfitting of any noisy data and in case of BioGP special procedures have been implemented to avoid bloat. Any pair of mutually conflicting objectives created through this procedure can also be optimized here using a built-in evolutionary strategy, incorporated as a module.