Article Preview
TopIntroduction
Artificial Neural Networks (ANNs) are the most novel and powerful artificial tool suitable for solving combinatorial problems such as prediction, forecasting and classification (Daqi & Yan, 2005; de A. Araújo, 2011; Ghazali, Jaafar Hussain, Mohd Nawi, & Mohamad, 2009). NNs are being used extensively for solving universal problems intelligently like continuous, discrete, telecommunications fraud detection and clustering (Charalampidis & Muldrey, 2009; Hilas & Mastorocostas, 2008; Mielniczuk & Tyrcha, 1993). Ns are being applied for different optimization and mathematical problems such as classification, object and image recognition, signal processing, temperature and weather forecasting and bankruptcy (Aussem, Murtagh, & Sarazin, 1994; Chaudhuri & Bhattacharya, 2000; Chen, Duan, Cai, & Liu, 2011; Uncini, 2003).
There are several techniques used for NNs favourable performance for training ANN such as evolutionary algorithms (EA), Multi-objective hybrid evolutionary algorithms (Qasem, Shamsuddin, & Zain, 2012; Tallón-Ballesteros & Hervás-Martínez, 2011), Genetic algorithms (GA) (Blanco, Delgado, & Pegalajar, 2001; García-Pedrajas, Ortiz-Boyer, & Hervás-Martínez, 2006; Geretti & Abramo, 2011), Particle swarm optimization (PSO) (Al-Shareef & Abbod, 2010; Hong-Bo, Yi-Yuan, Jun, & Ye, 2004; Hongwen & Rui, 2006), deferential evolution (DE) (Slowik & Bialko, 2008; Subudhi & Jena, 2011), an colony optimization (ACO) (Ashena & Moghadasi, 2011; Blum & Socha, 2005), BP d improved BP algorithm (Nawi, Ransing, & Ransing, 2006; Yan, Zhongjun, & Jiayu, 2010). These techniques are used for initialization of optimum weights, parameters, activation function, and selection of a proper network structure.
The main task of BP algorithm is to update the network weights for minimizing output error using BP processing because the accuracy of any approximation depends upon the selection of proper weights for the neural networks (NNs). It has high success rate in solving many complex problems, but it still has some drawbacks, especially when setting parameter values like initial values of connection weights, value for learning rate, and momentum. If the network topology is not carefully selected, the NNs algorithm can get trapped in local minima, or it might lead to slow convergence or even network failure. In order to overcome the disadvantages of standard BP, much global optimization population-based technique[20], GA (Geretti & Abramo, 2011), improved BP (Nawi, et al., 2006), DE (Slowik & Bialko, 2008), BP-ant colony(Chengzhi, Yifan, Lichao, & Yang, 2008), and PSO (Hongwen & Rui, 2006).