查看更多>>摘要：We consider a decentralized learning problem where training data samples are distributed over agents (processing nodes) of an underlying communication network topology without any central (master) node. Due to information privacy and security issues in a decentralized setup, nodes are not allowed to share their training data and only parameters of the neural network are allowed to be shared. This article investigates decentralized learning of randomization-based neural networks that provides centralized equivalent performance as if the full training data are available at a single node. We consider five randomization-based neural networks that use convex optimization for learning. Two of the five neural networks are shallow, and the others are deep. The use of convex optimization is the key to apply alternating-direction-method-of-multipliers with decentralized average consensus. This helps us to establish decentralized learning with centralized equivalence. For the underlying communication network topology, we use a doubly-stochastic network policy matrix and synchronous communications. Experiments with nine benchmark datasets show that the five neural networks provide good performance while requiring low computational and communication complexity for decentralized learning. The performance rankings of five neural networks using Friedman rank are also enclosed in the results, which are ELM < RVFL< dRVFL < edRVFL < SSFN. (c) 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
查看更多>>摘要：Large-scale optimization is the key to many practical decision processes. To deal with the dimensional issue in such problems, many approaches incorporate a divide-and-conquer strategy. Among them, cooperative coevolution approaches have recently gained popularity. Depending on the problem's structure, the decomposition of any large problem, into a number of smaller sub-problems, may leave some variables common in more than one sub-problem. Such a decomposition may have a negative effect on the quality of the final solution of an optimization problem. In this paper, we have proposed an algorithm that incorporates a novel decomposition method, where the objective of decomposition is to minimize the number of common variables between sub-problems, achieved by exploiting a variable interaction matrix developed from the problem. So the algorithm works as a two-stage approach, where the first stage is the problem decomposition, and the second stage is to find the solutions of the problem. The performance of our proposed algorithm is assessed by solving different sets of large-scale non-separable benchmark functions with up to 2,905 variables. The experimental results provide important insights into the efficiency of the proposed decomposition method, which in turn improves the performance of the optimization process. (C) 2021 Elsevier B.V. All rights reserved.
查看更多>>摘要：Predictive performance of Machine Learning (ML) models rely on the quality of data used for training the models. However, if the training data is not balanced among different classes, the performance of ML models deteriorate heavily. Several techniques have been proposed in the literature to add some semblance of balance to the data sets by adding artificial data points. Synthetic Minority Oversampling Technique(SMOTE) and Adaptive Synthetic Sampling(ADASYN) are some of the commonly used techniques to deal with class imbalance. However, these approaches are prone to 'within class imbalance' and 'small disjunct problem'. To overcome these problems, this article proposes an advanced algorithm by studying the compactness and location of the minority class relative to other classes. The proposed technique called K-Nearest Neighbor OveRsampling approach (KNNOR) performs a three step process to identify the critical and safe areas for augmentation and generate synthetic data points of the minority class. The relative density of the entire population is considered while generating artificial points. This enables the proposed KNNOR approach to oversample the minority class more reliably and at the same time stay resilient against noise. The proposed method is compared with the ten top performing contemporary oversamplers by testing the accuracy of classifiers trained on augmented data provided by each oversampler. The experimental results on several common imbalanced datasets show that our method ranks first more consistently than the other state-of-art oversamplers. The proposed method is easy to use and has been made open source as a python library. (c) 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
查看更多>>摘要：The financial situation of listed companies has a great impact on the construction of stock portfolio. However, some traditional portfolio models only consider the fluctuation of stock price and ignore the impact of the financial situation of listed companies on the portfolio, which will affect the effectiveness of the portfolio. To fill the gap, a new portfolio model based on the KMV model and a multiobjective water cycle algorithm is proposed to further improve the framework of portfolio construction. Firstly, the KMV model is used to evaluate the financial situation of listed companies, and then the optimization algorithm directly uses both the results of KMV and the fluctuation of stock price to build a more reasonable portfolio. To evaluate the validity of the model, the data of 100 A-share listed companies collected from China were used in two experiments. The results of the experiments show that the model can determine the stock portfolio according to the internal financial information and stock history information of listed companies, which not only can improve the effect and stability of the investment portfolio, but also can play a guiding role in the performance evaluation of listed companies. (C) 2021 Elsevier B.V. All rights reserved.
查看更多>>摘要：In data science there are problems that are not visible until you work with a sufficiently large number of data. This is the case, for example, with the design of the inference engine in fuzzy rule-based classification systems. The most common way to implement the winning rule inference method is to use sequential processing that reviews each of the rules in the rule set, to determine the best one and return the associated class. This implementation produces fast response times when the set of rules is small and is applied to a small set of examples. In this paper we explore new versions to implement this inference method, avoiding analyzing all the rules and focusing the analysis on the neighborhood of rules around the example. We study experimentally the conditions where each of them should be applied. Finally, we propose an implementation that combines all the studied versions offering good accuracy results and a significant reduction in the response time. (C) 2021 Elsevier B.V. All rights reserved.
查看更多>>摘要：During protein synthesis the genetic code links each codon, a triplet of nucleotides, with the corresponding amino acid. Synonymous codons are those that code for the same amino acid. The difference in the frequency of occurrence of certain synonymous codons over other synonymous codons is called the codon usage bias (CUB). The Zeng and Charlesworth model is used to estimate the strength of CUB. In their model the evolutionary process is represented by a Markov model, which allows the population size to vary over time. In this paper we propose a new method that incorporates demographic changes into the model. The method is a hybrid of two optimizers, the first is evolutionary programming and the second is a version of the genetic algorithms that uses chromosomes of variable lengths, which allows for expressing more demographic changes than what the simplified model presented by Zeng and Charlesworth does. We conduct several simulations to show why this hybridization is necessary, and also to show the superior performance of this new hybrid. (C) 2021 Elsevier B.V. All rights reserved.
查看更多>>摘要：Vortex detection methods help researchers to better understand the potential flow mechanism, and can be divided into three groups. Global methods have higher accuracy at the expense of time performance, while local methods provide results rapidly with poor accuracy. Machine learning-based methods consider both computational speed and accuracy, but their generalization and scalability are poor, which prevents them from being applied to real scenes. To address the above issues, we propose a novel vortex detection method, termed Vortex-U-Net. Our method has three characteristics. Firstly, our approach combines the characteristics of both global and local vortex detection methods. Secondly, it adopts the vorticity field, which integrates the velocity field and coordinates of grid points as the input. In this manner, it can keep more physical grid information of flow fields, which further improves the accuracy and generalization. Thirdly, our method fusions the properties of flow fields into the design of the loss function of the network. The proposed Vortex-U-Net model is subsequently evaluated against several widely used vortex detection methods on both numerically-simulated and analytical flows. Results reveal that our approach can achieve both high accuracy and performance. (C) 2021 Published by Elsevier B.V.
查看更多>>摘要：Time waveform characteristics of electromagnetic transients such as rise and fall time, duration or energy integral are vital criteria for reliability analysis, but it often meets the challenge to make an evaluation for the electronics of interest due to the lack of sufficient information from sampling and testing. Benefiting from the recent achievements on deep learning technique, in this paper a reliability assessment method for electronics is proposed based on neural network. The recurrent neural network (RNN) is involved to approximate the time waveform norm, so that the time-related characteristic can be extracted in metric space for comparison and classification. Inspired by the model-based few-shot learning strategy, a Siamese network architecture of two weights-sharing RNN is trained to avoid possible over-fitting. Artificial data representing various pulse waveforms are generated, with the help of which the approximation ability of RNN to two kinds of time waveform p-norms is analyzed and discussed in depth. To demonstrate the applicability of the proposed model, the lifetime stage of gas discharge tube (GDT) after cumulative discharge is experimentally investigated. The results are also compared with several informed evaluation model, and the proposed model is verified to able to yield the interpretable estimation in metric space and free from extra prior information. (c) 2021 Elsevier B.V. All rights reserved.
查看更多>>摘要：Graphs have gained a lot of attention in the pattern recognition community thanks to their ability to encode both topological and semantic information. Despite their invaluable descriptive power, their arbitrarily complex structured nature poses serious challenges when they are involved in learning systems. Typical approaches aim at building a vectorial representation of the graph in a suitable embedding space by leveraging on the selection of relevant prototypes that enable the use of common pattern recognition methods. An emerging paradigm able to synthesize prototypes in a data-driven fashion can be found in Granular Computing. Nonetheless, these methods often require a core dissimilarity measure defined directly in the graph domain that usually relies on a set of suitable parameters which are heavily problem-dependent. The automatic selection of these parameters is of utmost importance for building embedding spaces able to preserve the semantic contents between the structured and vector domains. In this paper, we propose an evolutionary-based approach for learning multiple dissimilarity measures tailored on each of the problem-related classes for the classification problem at hand. The learnt class-specific metrics contribute in synthesizing prototypes with high informative content related to each class by means of a Granular Computing approach. Such prototypes induce an embedding space where the graph classification can take place with common pattern recognition techniques for vector data. Tests conducted on publicly available datasets corroborate the effectiveness of the proposed approach both in terms of learning performances and interpretability of the model, as measured by the classification accuracy and number of meaningful prototypes considered in the synthesized model. (C) 2021 Elsevier B.V. All rights reserved.
查看更多>>摘要：In the modern power systems, appropriate power dispatch schedule of the online power generating units is essential for reliable and clean power supply and it is desirable to attain this at the lowest possible operating cost. Mathematical model of such an Economic Load Dispatch (ELD) problem turns out to be a complex non-convex optimization problem with practical constraints involving various factors such as, line loss, valve points effect, prohibited operating zones, ramp rate limits, system spinning reserve, multiple fuel options, etc. Nature inspired algorithms have been extensively used to solve such complex ELD problems. Through this study we propose to contribute to the available pool of efficient methodologies for solving ELD problems. Recently, Kumar et al. (2018) developed an ant colony optimization algorithm, namely ACO-LD which has shown impressive results compared to other efficient algorithms in literature when applied to a set of unconstrained optimization problems. In this study we extend ACO-LD to develop a new Constrained Ant Colony Optimization (ACO) algorithm with Adaptive Penalty (AP) method (Lemonge and Barbosa, 2014), to solve the ELD problem. The proposed algorithm is named CACO-LD-AP and is found to be quite efficient in terms of the quality of the solutions found for ELD problems of varied complexities. In order to validate the efficiency of the proposed algorithm, six power systems have been considered in this study. The performance of CACOLD-AP is compared with various recently published state of the art algorithms for solving ELD problems. Analysis of the experimental results affirms the robustness and superiority of CACO-LD-AP over other algorithms included in this study. (C) 2021 Elsevier B.V. All rights reserved.
NSTL
Elsevier