Search

Neural Architecture Search (NAS) is an automatic process of designing a neural architecture for solving classification problems. It is closely related to hyper-parameters such as hidden layers, neurons in each hidden layer, type of activation function (ACT), network optimizer and so on. Therefore, finding appropriate hyper-parameters to construct suitable network architecture for a particular problem is a challenging task. In this paper, an improved Genetic Algorithm (GA-NAS) is proposed to build a multi-layer feed forward architecture for image classification problem. Each chromosome of the proposed method is encoded with four hyper-parameters namely no. of hidden layers, neurons per hidden layer, activation function (ACT) and network error optimization technique. Each chromosome represents a neural network architecture for the given problem. The categorical cross-entropy or log function is considered to represent fitness function which provides performance accuracy of the architecture. The proposed methodology is experimented on two well-known benchmark image classification data sets such as CIFAR-10 and MNIST. The GA-NAS is compared with brute force algorithm and obtained results demonstrated the effectiveness for solving image classification problems.

Activation functions in the neural networks play an important role by introducing non-linear properties to the neural networks. Thus it is considered as one of the essential ingredients among other building blocks of a neural network. But the selection of the appropriate activation function for the enhancement of model accuracy is strenuous in a sense; the performance of the NN-model is influenced by a proper selection of activation function for a dataset. Proper activation function selection is still a trial and error method for which the model accuracy improves for classification. As a solution to this problem, we have proposed an activation function ensembling by majority voting that has significantly improved the model accuracy in a classification context. The proposed model is tested on four benchmark datasets such as MNIST, Fashion MNIST, Semeion, and ARDIS IV datasets. The result shows that the performance of the proposed model is appreciably better than other traditional methods such as Convolutional Neural Network (CNN), Support Vector Machine (SVM), Recurrent Neural Network (RNN).

Drug is a very much essential substance in health care system. Producing a new drug for a disease in the market using traditional method is very time consuming and expensive. Recently, drug design process is sped-up by using computer resources known as Computer-Aided Drug Design (CADD). In drug design, bio-molecules are responsible to produce a new drug. Therefore, molecules identification is an essential part of CADD. In this paper, two machine learning algorithms such as Naive Bayesian (NB) classifier and k-Nearest Neighbors (k-NN) are evaluated to classify multi-class drug molecules. The result reveals that k-NN method shows higher Accuracy and higher Precision compare to NB. Furthermore, Recall and F1-score of k-NN are higher than that of NB.

Traditional drug designing and discovery processes are a very time consuming, expensive and challenging tasks to produce a new drug in the market. Computer Aided Drug Design (CADD) is a promising approach that is cost effective as well as speeds up the drug designing process. CADD is a computational methods which provides resources for simplifying the design and discovery of a new drug. At molecular level, a drug binds to the target protein and neutralizes the disease. Therefore, identification of active molecules which can bind to the target protein is an essential part of CADD. In this paper, the back propagation neural network model is employed for predicting inactive or active molecules, which provides chemical compounds with desirable properties for drug design. The proposed approach demonstrates 99% prediction accuracy on the dataset which consists of active and inactive molecules taken from PubChem data repository.

Artificial Neural Networks (ANNs) have been successfully applied to solve various classification problems. To perform the classification task,neural network training is essential. Multiple parameters are associated with neural network training and setting these parameters are a very challenging task. Several meta-heuristic algorithms are employed to train the network. In this paper, Particle Swarm optimization (PSO) is used to train the network to optimize the training error while solving a classification problem. Performance of the proposed method is evaluated on ten different data sets which are in supervised. The experimental result shows a significant performance of the proposed methodology compared to traditional neural network training approaches.

In the field of Image Processing, Image segmentation is a low level but important task in entire image understanding system which divides an image into its multiple disjoint regions based on homogeneity. In most of the machine vesion and high level image understanding application this is one of the important steps. Till date different techniques of image segmentation are available and hence There exists a huge survey literature in different approaches of Image Segmentation. Selection of image segmentation technique is highly problem specific. There is no versatile algorithm which is applicable for all kinds of images. Optimization based image segmentation is not explored much which can be applied to reduce complexity of the problem. The aim of the paper is to search for an optimized threshold value for Image Segmentation using Particle Swarm Optimization (PSO) algorithm where fitness function is designed based on entropy of the image.

In recent decade Evolutionary Algorithms plays an important role in many engineering design and optimization problems. Particle Swarm Optimization (PSO) is one of such algorithm which is based on the intelligent food searching behavior of swarm like birds flock, fish schooling. It has been shown that it works efficiently on noisy, multimodal and composite functions. However, it stuck at local optima at later stage of evolution due to unexplore the search space. Several variations of pso and mutation based approached was developed for this problem. In this paper, an adaptive mutation is proposed for multiobjective pso and called it AMPSO. In AMPSO, mutation is applied on the position and velocity of the particles based on the fitness values of the particles. Proposed algorithm carried on 5 multiobjective benchmark functions. The experimental results shown the better performance comparing with other algorithms in terms of best, mean and standard deviation.

The Particle Swarm Optimization (PSO) algorithm is a novel population based swarm algorithm has shown good performance on well-known numerical test problems. However, PSO tends to suffer from premature convergence on multimodal test problems. This is due to lack of diversity of population in search space and leads to stuck at local optima and ultimately fitness stagnation of the population. To enhance the performance of PSO algorithms, in this paper, we propose a method of population adaptation (PA). The proposed method can identify the moment when the population diversity is poor or the population stagnates by measuring the Euclidean distance between particle position and particles average position of a population. When stagnation in the population is identified, the population will be regenerated by normal distribution to increase diversity in the population. The population adaptation is incorporated into the PSO algorithm and is tested on a set of 13 scalable CEC05 benchmark functions. The results show that the proposed population adaptation algorithm can significantly improve the performance of the PSO algorithm with standard PSO, ATREPSO and ARPSO.

Artificial Bee Colony (ABC) is a swarm based stochastic search algorithm inspired by the foraging behavior of honeybees. Due to the simplicity of implementation and promising optimization capability, ABC is successfully applied to solve wide class of scientific and engineering optimization problems. But, it has problems of premature convergence and trapping in local optima. In this paper, to enhance the performance of ABC, we have proposed a modified version of ABC algorithm using Differential Evolution (DE) and Polynomial Mutation (PM) called DE-PM-ABC. The comparison with ABC by Karaboga [1], MABC [27] by Liu et al. using some benchmark functions of CEC 2005 demonstrates that our approach achieves a good trade-off between exploration and exploitation and thus obtains better global optimization result and faster convergence speed.

Several population based search algorithms are developed by the researchers to predict the native state of protein from its primary sequences. The paper aims at predicting the native conformation of proteins in lattice model using PSO based searching method. However, stuck at local minima and generating illegal conformation are the main drawbacks of applying the search algorithm in protein structure prediction. Adaptive Polynomial Mutation (APM) is performed to remove local minima while illegal conformations are repaired using backtracking method. Benchmark sequences with different length are applied to verify the proposed algorithm showing better results compare to the earlier approaches.

Particle Swarm Optimization (PSO) has shown its good search ability in many optimization problem. But PSO easily gets trapped into local optima while dealing with complex problems. In this work, we proposed an improved PSO, namely PSO-APM, in which adaptive polynomial mutation strategy is employed on global best particle with the hope that it will help the particles jump out local optima. In this work, we carried out our experiments on 8 well-known benchmark problems. Finally the results are compared with classical PSO and PSO with power mutation (PMPSO).