Search

WOS: 000416498700046, In this study, the cyclic variability of a diesel engine using diesel fuel and butanol diesel fuel blends is modeled using an artificial neural network (ANN) method. The engine was operated with ten different engine speeds and full load conditions using six different n-butanol diesel fuel blends. The coefficient of variation (COV) of the indicated mean effective pressure (IMEP), which is a well-accepted evaluation method, was used to assess the cyclic variability for 100 sequential engine cycles. Results indicated that adding n-butanol to diesel fuel caused an increase. Moreover, the COVimep values exhibited a decreasing trend with an increase in the engine speed for each fuel. The experimental results were used to train the ANN. The ANN network was trained with Levenberg - Marquardt (LM) and Scaled Conjugate Gradient (SCG) algorithms. After training the ANN, it was found that the coefficient of determination (R-2) values were in the range of between 0.737 and 0.9677, the mean-absolute-percentage error (MAPE) values were smaller than 8.7 and the mean-square error values (MSE) were smaller than 0.042. The predictions of the developed ANN model showed reasonable consistency with the experimental results. (C) 2017 Elsevier Ltd. All rights reserved., Research Fund of Karadeniz Technical University [FYL-2015-5286], This work was supported by the Research Fund of Karadeniz Technical University, Project number: FYL-2015-5286.

Predictions on sequential data, when both the upstream and downstream information is important, is a difficult and challenging task. The Bidirectional Recurrent Neural Network (BRNN) architecture has been designed to deal with this class of problems. In this paper, we present the development and implementation of the Scaled Conjugate Gradient (SCG) learning algorithm for BRNN architectures. The model has been tested on the Protein Secondary Structure Prediction (PSSP) and Transmembrane Protein Topology Prediction problems (TMPTP). Our method currently achieves preliminary results close to 73% correct predictions for the PSSP problem and close to 79% for the TMPTP problem, which are expected to increase with larger datasets, external rules, ensemble methods and filtering techniques. Importantly, the SCG algorithm is training the BRNN architecture approximately 3 times faster than the Backpropagation Through Time (BPTT) algorithm. © Springer International Publishing Switzerland 2016. 9886 LNCS 123 131 Sponsors: Conference code: 180929

The model of multi-layered neural networks of the back-propagation type is well-known for their universal approximation capability and even the standard back-propagation training algorithm used for their adjustment often provides results applicable to real-world problems. The present study deals with the issue of the multi-layered neural networks. It describes selected variants of training algorithms, mainly the standard back-propagation training algorithm and the scaled conjugate gradients algorithm, which ranks among the extremely fast second-order algorithms. One of the parts of the present study is also an application for the visualisation of the structure of multi-layered neural networks whose solution is designed with respect to its potential utilization in the education of artificial intelligence. The first part of the study introduces the subject matter and formally describes both algorithms, followed by a short description of other variants of the algorithms and their analysis. The next part discusses the selection of the appropriate programming language for the implementation of the application, specifies the goals and describes the implementation works. The conclusion summarizes the test results of the speed and implementation comparison with the selected noncommercial-based software ENCOG.

Multi-layered neural networks of the back-propagation type are well known for their universal approximation capability. Already the standard back-propagation training algorithm used for their adjustment provides often applicable results. However, e cient solutions to complex tasks currently dealt with require a quick convergence and a transparent network structure. This supports both an improved generalization capability of the formed networks and an easier interpretation of their function later on. Various techniques used to optimize the structure of the networks like learning with hints; pruning and sensitivity analysis are expected to impact a better generalization, too. One of the fast learning algorithms is the conjugate gradient method. In this thesis, we discuss, test and analyze the above-mentioned methods. Then, we derive a new technique combining together the advantages of them. The proposed algorithm is based on the rapid scaled conjugate gradient technique. This classical method is enhanced with the enforcement of a transparent internal knowledge representation and with pruning of redundant hidden and input neurons inspired by sensitivity analysis. The performance of the developed technique has been tested on arti cial data and on real-world data obtained from the World Bank. Experiments done...