Quick flower pollination algorithm (QFPA) and its performance on neural network training

This study proposes a novel version of flower pollination algorithm (FPA) and it is called as quick flower pollination algorithm (QFPA). Two important changes are carried out to improve local and global search capability in QFPA. Firstly, switch probability is determined according to number of generations adaptively, unlike standard FPA. Secondly, solution generation mechanism used in local pollination is updated by arithmetic crossover. Two different problem groups are utilized to evaluate the performance of QFPA: solution of global optimization problems and training of artificial neural network. Firstly, 56 benchmark test functions are used for analysis of global optimization problems. Secondly, artificial neural network is trained by QFPA to identify nonlinear dynamic systems and four different nonlinear dynamic systems are utilized. Root-mean-square error (RMSE) is choosen as the performance metric. In the identification of nonlinear systems based on neural network, QFPA provides up to 60% performance improvement compared to standard FPA. The results obtained for both problem types are compared with bee algorithm, harmonic search, artificial bee colony algorithm, standard FPA and some variants of FPA. The Wilcoxon signed rank test is used to determine significance of the results belonging to neural network training. The results show that QFPA is generally more effective than related meta-heuristic algorithms in both problem types.