An improved artificial bee colony algorithm for training multilayer perceptron in time series prediction

• Main Author: Shah, Habib
• Format: Thesis
• Language: English; English
• Published: 2014
• Subjects: QA76 Computer software; QA71-90 Instruments and machines
• Online Access: http://eprints.uthm.edu.my/1214/1/24p%20HABIB%20SHAH.pdf; http://eprints.uthm.edu.my/1214/2/HABIB%20SHAH%20WATERMARK.pdf; http://eprints.uthm.edu.my/1214/

Learning an Artificial Neural Network (ANN) is an optimization task since it is desirable to find optimal weight sets of an ANN in the training process. Different equations are used to guide the network for providing an accurate result with less training and testing error. Most of the training algorithms focus on weight values, activation functions, and network structures for providing optimal outputs. Backpropagation (BP) learning algorithm is the well-known learning technique that trained ANN. However, some difficulties arise where the BP cannot get achievements without trapping in local minima and converge very slow in the solution space. Therefore, to overcome the trapping difficulties, slow convergence and difficulties in finding optimal weight values, three improved Artificial Bee Colony (ABC) algorithms built on the social insect behavior are proposed in this research for training ANN, namely the widely used Multilayer Perceptron (MLP). Here, three improved learning approaches inspired by artificial honey bee's behavior are used to train MLP. They are: Global Guided Artificial Bee Colony (GGABC), Improved Gbest Guided Artificial Bee Colony (IGGABC) and Artificial Smart Bee Colony (ASBC) algorithm. These improved algorithms were used to increase the exploration, exploitation and keep them balance for getting optimal results for a given task. Furthermore, here these algorithms used to train the MLP on two tasks; the seismic event's prediction and Boolean function classification. The simulation results of the MLP trained with improved algorithms were compared with that when trained with the standard BP, ABC, Global ABC and Particle Swarm Optimization algorithm. From the experimental analysis, the proposed improved algorithms get better the classification efficacy for time series prediction and Boolean function classification. Moreover, these improved algorithm's success to get high accuracy and optimize the best network's weight values for training the MLP.