Regression modeling based on improved genetic algorithm

Abstract

Original scientific paper Regression model is a well-established method in data analysis with applications in various fields. The selection of independent variables and mathematically transformed in a regression model is often a challenging problem. Recently, some scholars have used evolutionary computation to solve this problem, but the result is not effective as we desired. The crossover operation in GA is redesigned by using Latin hypercube sampling, then combining two commonly used statistical criteria (AIC, BIC) we are presenting an improved genetic algorithm based for solving statistical model selection problem. The proposed algorithm can overcome strong path-dependence and rely on experience of classical approaches. Comparison of simulation results in solving statistical model selection problem with this improved GA, traditional genetic algorithm and classical algorithm for model selection show that the new GA has superiority in solution of quality, convergence rate and other various indices. Keywords: genetic algorithm; Latin hypercube sampling; regression analysis; regression model selection Regresijsko modeliranje zasnovano na poboljšanom genetičkom algoritmuIzvorni znanstveni članak Regresijski model je dobro uhodana metoda u analizi podataka s primjenom u raznim područjima. Izbor nezavisnih varijabli i matematički transformiranih u regresijski model, često predstavlja izazovan problem. Nedavno je nekoliko znanstvenika primijenilo evolucijski proračun za rješenje tog problema, ali rezultat nije učinkovit onoliko koliko smo željeli. Ukrižena (crossover) operacija u GA redizajnirana je primjenom Latin hypercube uzorkovanja, a zatim, kombinacijom dvaju uobičajeno korištenih statističkih kriterija (AIC, BIC), dajemo poboljšani genetički algoritam za rješavanje problema izbora statističkog modela. Predloženim se algoritmom može prevladati jaka ovisnost o putanji i osloniti na iskustvo stečeno primjenom klasičnih pristupa. Usporedba rezultata simulacije u rješavanju problema odabira statističkog modela s ovim poboljšanim GA, tradicionalnog genetičkog algoritma i klasičnog algoritma za odabir modela pokazuje da je novi GA superiorniji u rješavanju kvalitete, brzine konvergencije i drugih različitih pokazatelja.

A QSAR modeling approach for predicting myeloid antimicrobial peptides with high sequence similarity

“…The reduced pool of descriptors was scaled, and models were built using genetic algorithm (GA) optimization technique which aims to find the best combination of variables without exploring all of them. As GA is self-adaptive and self-learning, it is highly efficient at searching a subset of features to find optimal combinations, especially when the search space is large (Johnson et al, 2014;Minghua, Qingxian, Benda, & Feng, 2017). GA was coupled with the MLR fitness evaluator in the QSARINS program (QSAR-INSUBRIA; www.qsar.it; Gramatica, Cassani, & Chirico, 2014;Gramatica, Chirico, Papa, Cassani, & Kovarich, 2013).…”