Neural Computing and Applications

Sun, Yanan; Mao, Hua; Sang, Yongsheng; Yi, Zhang

doi: 10.1007/s00521-015-2082-xpmid: N/A

The auto-encoder model plays a crucial role in the success of deep learning. During the pre-training phase, auto-encoders learn a representation that helps improve the performance of the entire neural network during the fine-tuning phase of deep learning. However, the learned representation is not always meaningful and the network does not necessarily achieve higher performance with such representation because auto-encoders are trained in an unsupervised manner without knowing the specific task targeted in the fine-tuning phase. In this paper, we propose a novel approach to train auto-encoders by adding an explicit guiding term to the traditional reconstruction cost function that encourages the auto-encoder to learn meaningful features. Particularly, the guiding term is the classification error with respect to the representation learned by the auto-encoder, and a meaningful representation means that a network using the representation as input has a low classification error in a classification task. In our experiments, we show that the additional explicit guiding term helps the auto-encoder understand the prospective target in advance. During learning, it can drive the learning toward a minimum with better generalization with respect to the particular supervised task on the dataset. Over a range of image classification benchmarks, we achieve equal or superior results to baseline auto-encoders with the same configuration.

Kerh, Tienfuan; Su, Yu-Hsiang; Mosallam, Ayman

doi: 10.1007/s00521-015-2077-7pmid: N/A

In this study, a genetic algorithm with global searching capability was incorporated into a neural network calculating process to obtain a highly reliable model for predicting peak ground acceleration, which is the key element in evaluating earthquake response and in establishing a seismic design standard. In addition to three seismic parameters (i.e. local magnitude, focal distance, and epicentre depth), this study included two geological conditions (i.e. standard penetration test value and shear-wave velocity) in the input to reflect the site response adequately. Based on the earthquake records and soil test data from 86 checking stations, within 24 seismic subdivision zones in the Taiwan area, the computational results show that using a combination of a neural network and genetic algorithm can achieve a higher performance compared with solely using a neural network model. Furthermore, a weight-based model was developed for predicting peak ground acceleration at an unmonitored site to represent each subdivision zone. The results show that three subdivision zones have higher horizontal peak ground accelerations than the seismic design value as required in the building code. The obtained information might be helpful in relevant engineering applications for the studied region, and the proposed method for treating this type of nonlinear seismic data might be applicable in other areas of interest worldwide.

Rakkiyappan, R.; Dharani, S.

doi: 10.1007/s00521-015-2079-5pmid: N/A

This paper is devoted to investigate the problem of global asymptotic synchronization of an array of N randomly coupled reaction–diffusion neural networks with Markovian jumping parameters and mixed delays using sampled-data control technique. The jump parameters are determined by a continuous-time, discrete-state Markovian chain, and the mixed time delays under consideration comprise both discrete and distributed delays. A multiple integral inequality is proposed firstly in Markovian jump reaction–diffusion neural networks with mixed delays. Through constructing appropriate Lyapunov–Krasovskii functional including multiple integral terms, some novel synchronization criteria in terms of linear matrix inequalities are derived. The obtained LMIs can be easily verified for feasibility through any of the available softwares. Finally, numerical examples with simulations are provided to illustrate the effectiveness of the proposed theoretical results.

Awadallah, Mohammed; Al-Betar, Mohammed; Khader, Ahamad; Bolaji, Asaju; Alkoffash, Mahmud

doi: 10.1007/s00521-015-2076-8pmid: N/A

This paper proposes a hybrid harmony search algorithm (HHSA) for solving the highly constrained nurse rostering problem (NRP). The NRP is a combinatorial optimization problem tackled by assigning a set of shifts to a set of nurses; each has specific skills and work contract, to a predefined rostering period according to a set of constraints. The harmony search is a metaheuristic approach, where the metaheuristics are the most successful methods for tackling this problem. In HHSA, the harmony search algorithm is hybridized with the hill climbing optimizer to empower its exploitation capability. Furthermore, the memory consideration operator of the HHSA is modified by replacing the random selection scheme with the global-best concept of particle swarm optimization to accelerate its convergence rate. The standard dataset published in the first international nurse rostering competition 2010 (INRC2010) was utilized to evaluate the proposed HHSA. Several convergence scenarios have been employed to study the effects of the two HHSA modifications. Finally, a comparative evaluation against twelve other methods that worked on the INRC2010 dataset is carried out. The experimental results show that the proposed method achieved five new best results, and 33 best published results out of 69 instances as achieved by other comparative methods.

Wu, Fei; Jing, Xiao-Yuan; Liu, Qian; Wu, Song-Song; He, Guo-Liang

doi: 10.1007/s00521-015-2081-ypmid: N/A

Kernel discriminant subspace learning technique is effective to exploit the structure of image dataset in the high-dimensional nonlinear space. However, for large-scale image recognition applications, this technique usually suffers from large computational burden. Although some kernel accelerating methods have been presented, how to greatly reduce computing time and simultaneously keep favorable recognition accuracy is still challenging. In this paper, we introduce the idea of parallel computing into kernel subspace learning and build a parallel kernel discriminant subspace learning framework. In this framework, we firstly design a random non-overlapping equal data division strategy to divide the whole training set into several subsets and assign each computational node a subset. Then, we separately learn kernel discriminant subspaces from these subsets without mutual communications and finally select the most appropriate subspace to classify test samples. Under the built framework, we propose two novel kernel subspace learning approaches, i.e., parallel kernel discriminant analysis (PKDA) and parallel kernel semi-supervised discriminant analysis (PKSDA). We show the superiority of the proposed approaches in terms of time complexity as compared with related methods, and provide the fundamental supports for our framework. For experiment, we establish a parallel computing environment and employ three public large-scale image databases as experiment data. Experimental results demonstrate the efficiency and effectiveness of the proposed approaches.

Fallahpour, Alireza; Olugu, Ezutah; Musa, Siti

doi: 10.1007/s00521-015-2078-6pmid: N/A

Supplier evaluation and selection is a complicated process which deals with conflicting attributes such as quality, cost. To mitigate the computational complexity, intelligent-based techniques have gained much popularity. But the main shortcoming of the existing models in this regard is to be a black box system. In this paper, we aim to combine analytical hierarchy process with multi-expression programming to both introduce a new evolutionary approach in the field of supplier evaluation and selection and cope with the earlier problem. To show the validity of the model, statistical test was carried out. The finding showed that the proposed model is accurate and acceptable for using in the evaluation process.

Demertzis, Kostantinos; Iliadis, Lazaros; Avramidis, Stavros; El-Kassaby, Yousry

doi: 10.1007/s00521-015-2075-9pmid: N/A

Several machine learning models were used to predict interior spruce wood density using data from open-pollinated progeny testing trial. The data set consists of growth (height and diameter which were used to estimate individual tree volume) and wood quality (wood density determined by X-ray densitometry, resistance to drilling, and acoustic velocity) attributes for a total of 1146 trees growing on comparable sites in interior British Columbia. Various machine learning models were developed for estimating wood density. The multilayer feed-forward artificial neural networks and gene expression programming provided the highest predictability as compared to the other methods tested, including those based on classical multiple regression which was considered as the comparisons benchmark. The utilization of machine learning models as a credible method for estimating wood density using available growth data as an indirect method for determining trees wood density is expected to become increasingly helpful to forest managers and tree breeders.

Liu, Yi-Cheng; Yeh, I-Cheng

doi: 10.1007/s00521-015-2090-xpmid: N/A

There are three disadvantages of weighted scoring stock selection models. First, they cannot identify the relations between weights of stock-picking concepts and performances of portfolios. Second, they cannot systematically discover the optimal combination for weights of concepts to optimize the performances. Third, they are unable to meet various investors’ preferences. This study aimed to more efficiently construct weighted scoring stock selection models to overcome these disadvantages. Since the weights of stock-picking concepts in a weighted scoring stock selection model can be regarded as components in a mixture, we used the simplex-centroid mixture design to obtain the experimental sets of weights. These sets of weights are simulated with US stock market historical data to obtain their performances. Performance prediction models were built with the simulated performance data set and artificial neural networks. Furthermore, the optimization models to reflect investors’ preferences were built up, and the performance prediction models were employed as the kernel of the optimization models, so that the optimal solutions can now be solved with optimization techniques. The empirical values of the performances of the optimal weighting combinations generated by the optimization models showed that they can meet various investors’ preferences and outperform those of S&P’s 500 not only during the training period but also during the testing period.

kumar, Yugal; Sahoo, G.

doi: 10.1007/s00521-015-2095-5pmid: N/A

In the field of data analysis, clustering is a powerful technique which groups the data into different subsets using a distance function. Data belonging to the same subset are similar in nature and offer heterogeneity to the data that reside in other subsets. Clustering has proved its potentiality in various fields such as bioinformatics, pattern recognition, image processing and many more. In this paper, a two-step artificial bee colony (ABC) algorithm is proposed for efficient data clustering. In two-step ABC algorithm, the initial positions of food sources are identified using the K-means algorithm instead of random initialization. Along this, to discover the promising search areas, an improved solution search equation based on social behavior of PSO is applied in the onlooker bee phase of ABC algorithm and abandoned food source location is found by using Hooke and Jeeves-based direct search method. Five benchmark and two artificial datasets are applied to validate the proposed modifications in the ABC algorithm, and results of this study are compared with other well-known clustering algorithms. Both the experimental and statistical analyses show that improvements in ABC algorithm have an advantage over the conventional ABC algorithm for solving clustering problems.

Bhattacharyay, Dipankar; Kocaefe, Duygu; Kocaefe, Yasar; Morais, Brigitte

doi: 10.1007/s00521-015-2093-7pmid: N/A

Carbon anode is one of the key components for the electrolytic production of aluminum. It is mainly composed of calcined petroleum coke, coal tar pitch, and recycled carbon materials. The impurities in the raw materials, which are mainly by-products of different industries, influence significantly the quality of anodes. Usually, no well-known mathematical relationship exists between the various physical and chemical properties of raw materials and the final anode properties. In such situations, the artificial neural network (ANN) methods can serve as a useful tool to predict anode properties. In this study, published data have been used to show the proficiency of different artificial neural networks using the MATLAB software. The average error between the predicted and experimental values is around 6 %. The artificial neural network was also used to identify the effect of impurities such as, vanadium, iron, sodium, and sulfur on the CO2 reactivity of anodes. ANN also showed the effect of pitch percentage and coke porosity on the CO2 reactivity of anodes. The effect of CO2 and air reactivities of coke on the CO2 reactivity of anode was also studied. The predictions were found to be in good agreement with the results of other studies in the literature.