An improved bat algorithm with artificial neural networks for classification problems

Rehman Gillani, Syed Muhammad Zubair (2016) An improved bat algorithm with artificial neural networks for classification problems. PhD thesis, Universiti Tun Hussein Onn Malaysia.



Metaheuristic search algorithms have been used for quite a while to optimally solve complex searching problems with ease. Nowadays, nature inspired swarm intelligent algorithms have become quite popular due to their propensity for finding optimal solutions with agility. Moreover several algorithms belonging to the stochastic and deterministic classes are available (i.e. ABC, HS, CS, WS, BPNN, LM, and ERNN etc.). Recently, a new metaheuristic search Bat algorithm has become quite popular due its tendency towards convergence to optimal points in the search trajectory by using echo-location behavior of bats as its random walk. However, Bat suffers from large step lengths that sometimes make it to converge to sub-optimal solution. Therefore, in order to improve the exploration and exploitation behavior of bats, this research proposed an improved Bat with Gaussian Distribution (BAGD) algorithm that takes small step lengths and ensures convergence to global optima. Then, the proposed BAGD algorithm is further hybridized with Simulated Annealing (SA) and Genetic Algorithm (GA) to perform two stage optimization in which the former algorithm finds the optimal solution and the latter algorithm starts from where the first one is converged. This multi-stage optimization ensures that optimal solution is always reached. The proposed BAGD, SABa, and GBa are tested on several benchmark functions and improvements in convergence to global optima were detected. Finally in this research, the proposed BAGD, SABa, and GBa are used to enhance the convergence properties of BPNN, LM, and ERNN with proper estimation of the initial weights. The proposed Bat variants with ANN such as; Bat-BP, BALM, BAGD-LM, BAGD-RNN, GBa-LM, GBa-RNN, SABa-RNN, and SABa-LM are evaluated and compared with ABC-BP, and ABC-LM algorithms on seven benchmark datasets. From the simulation results, it can be realized that the proposed Bat algorithms with ANN outperforms the other algorithms in terms of CPU time, Mean Squared Error (MSE), and accuracy during convergence to global minima.

Item Type:Thesis (PhD)
Subjects:Q Science > QA Mathematics > QA76 Computer software
ID Code:9127
Deposited By:Mr. Mohammad Shaifulrip Ithnin
Deposited On:24 May 2017 15:58
Last Modified:24 May 2017 15:58

Repository Staff Only: item control page