Search

In this paper, the learning process of multilayer perceptron (MLP) neural network is boosted using hybrid metaheuristic optimization algorithms. Normally, the learning process in MLP requires suitable settings of its weight and bias parameters. In the original version of MLP, the gradient descent algorithm is used as a learner in MLP which suffers from two chronic problems: local minima and slow convergence. In this paper, six versions of memetic algorithms (MAs) are proposed to replace gradient descent learning mechanism of MLP where adaptive β -hill climbing (A β HC) as a local search algorithm is hybridized with six population-based metaheuristics which are hybrid flower pollination algorithm, hybrid salp swarm algorithm, hybrid crow search algorithm, hybrid grey wolf optimization (HGWO), hybrid particle swarm optimization, and hybrid JAYA algorithm. This is to show the effect of the proposed MA versions on the performance of MLP. To evaluate the proposed MA versions for MLP, 15 classification benchmark problems with different size and complexity are used. The A β HC algorithm is invoked in the improvement loop of any MA version with a probability of B r parameter, which is investigated to monitor its effect on the behavior of the proposed MA versions. The B r setting which obtains the most promising results is then used to set the hybrid MA. The results show that the proposed MA versions excel the original algorithms. Moreover, HGWO outperforms all other MA versions in almost all the datasets. In a nutshell, MAs are a good choice for training MLP to produce results with high accuracy. [ABSTRACT FROM AUTHOR]

High-dimensional Feature Selection Problems (HFSPs) have grown in popularity but remain challenging. When faced with such complex situations, the majority of currently employed Feature Selection (FS) methods for these problems drastically underperform in terms of effectiveness. To address HFSPs, a new Binary variant of the Ali Baba and the Forty Thieves (BAFT) algorithm known as binary adaptive elite opposition-based AFT (BAEOAFT), incorporating historical information and dimensional mutation is presented. The entire population is dynamically separated into two subpopulations in order to maintain population variety, and information and knowledge about individuals are extracted to offer adaptive and dynamic strategies in both subpopulations. Based on the individuals’ history knowledge, Adaptive Tracking Distance (ATD) and Adaptive Perceptive Possibility (APP) schemes are presented for the exploration and exploitation subpopulations. A dynamic dimension mutation technique is used in the exploration subpopulation to enhance BAEOAFT’s capacity in solving HFSPs. Meanwhile, the exploratory subpopulation uses Dlite Dynamic opposite Learning (EDL) to promote individual variety. Even if the exploitation group prematurely converges, the exploration subpopulation’s variety can still be preserved. The proposed BAEOAFT-based FS technique was assessed by utilizing the k-nearest neighbor classifier on 20 HFSPs obtained from the UCI repository. The developed BAEOAFT achieved classification accuracy rates greater than those of its competitors and the conventional BAFT in more than 90% of the applied datasets. Additionally, BAEOAFT outperformed its rivals in terms of reduction rates while selecting the fewest number of features. [ABSTRACT FROM AUTHOR]