关键词： Formal verification   classification ML   algorithm selection  

摘要： Typical modem HW designs include many blocks associated with thousands of design properties. Having today's commercial formal verifiers utilize a complementary set of state-of-art formal algorithms is a key in enabling the formal verification tools to successfully cope with verification problems of different sizes, types, and complexities. Formal engines orchestration is the methodology used to pick the most appropriate formal engine for a specific verification problem. It assures proper scheduling of the formal engines to minimize the time consumed to solve individual design verification problems, hence highly impacts the time required to verify the overall design properties. This work proposes the utilization of supervised machine learning classification techniques to guide the orchestration step by predicting the formal engines that should be assigned to a design property. Up to 16,500 formal verification runs on RTL designs and their properties are used to train the classifier to create a prediction model. The classifier assigns any new verification problem to an appropriate list of formal engines associated with a probability distribution over the set of engines classes. Our results indicate how the proposed model is able to improve the formal suite total run-time by up to 59% of its maximum allowable time improvement using multi-classification-based orchestration and to nominate with 88% accuracy the appropriate formal engines for new-to-verify HW designs.

关键词： PM2.5   Aerosol optical depth   Random forests   Machine learning   China  

摘要： Background: Machine learning algorithms have very high predictive ability. However, no study has used machine learning to estimate historical concentrations of PM2.5 (particulate matter with aerodynamic diameter <= 2.5 mu m) at daily time scale in China at a national level. Objectives: To estimate daily concentrations of PM2.5 across China during 2005-2016. Methods: Daily ground-level PM2.5 data were obtained from 1479 stations across China during 2014-2016. Data on aerosol optical depth (AOD), meteorological conditions and other predictors were downloaded. A random forests model (non-parametric machine learning algorithms) and two traditional regression models were developed to estimate ground-level PM2.5 concentrations. The best-fit model was then utilized to estimate the daily concentrations of PM2.5 across China with a resolution of 0.1 degrees (approximate to 10 km) during 2005-2016. Results: The daily random forests model showed much higher predictive accuracy than the other two traditional regression models, explaining the majority of spatial variability in daily PM2.5 [10-fold cross-validation (CV) R-2 = 83%, root mean squared prediction error (RMSE) = 28.1 mu g/m(3)]. At the monthly and annual time-scale, the explained variability of average PM2.5 increased up to 86% (RMSE = 10.7 mu g/m(3) and 6.9 mu g/m(3), respectively). Conclusions: Taking advantage of a novel application of modeling framework and the most recent ground-level PM2.5 observations, the machine learning method showed higher predictive ability than previous studies. Capsule: Random forests approach can be used to estimate historical exposure to PM2.5 in China with high accuracy. (C) 2018 Elsevier B.V. All rights reserved.

关键词： EGFR抑制剂   机器学习方法   随机森林   分子描述符  

摘要： 表皮生长因子受体(Epidermal Growth Factor Receptor,EGFR)的扩增或突变与人类实体肿瘤密切相关。三种机器学习方法:随机森林(random forest,RF),k最近邻和C4.5决策树,被用于建立理论模型,来预测EGFR的抑制剂。结果表明,这些模型的预测结果均取得了令人满意的精度。又通过对比分析,发现RF模型具有较好的预测性能,并进一步修改和优化了RF模型的参数。此外,还采用特征选择程序筛选出了25个和EGFR的抑制最相关的分子描述符。最后对得到的最优RF模型进行了评价。

关键词： CRISPR/Cas9   机器学习   sgRNA   脱靶效应   基因敲除  

摘要： 基于CRISPR/Cas9系统介导的第三代基因组定点编辑技术,已被广泛应用于基因编辑和基因表达调控等研究领域。如何提高该技术对基因组编辑的效率与特异性、最大限度降低脱靶风险一直是该领域的难点。近年来,机器学习为解决CRISPR/Cas9系统所面临的问题提供了新思路,基于机器学习的CRISPR/Cas9系统已逐渐成为研究热点。本文阐述了CRISPR/Cas9的作用机理,总结了现阶段该技术面临的基因组编辑效率低、存在潜在的脱靶效应、前间区序列邻近基序(PAM)限制识别序列等问题,最后对机器学习应用于优化设计高效向导RNA(sg RNA)序列、预测sg RNA的活性、脱靶效应评估、基因敲除、高通量功能基因筛选等领域的研究现状与发展前景进行了展望,以期为基因组编辑领域的研究提供参考。

关键词： Building energy   Data-driven   Machine learning   Multi-scale   Simulation   Urban energy modeling   Urban context  

摘要： The world is rapidly urbanizing, and the energy intensive built environment is becoming increasingly responsible for the world's energy consumption and associated environmental emissions. As a result, significant efforts have been put forth to develop methods that can accurately model and characterize building energy consumption in cities. These models aim to utilize physics-based building energy simulations, reduced-order calculations and statistical learning methods to assess the energy performance of buildings within a dense urban area. However, current urban building energy models are limited in their ability to account for the inter-building energy dynamics and urban microclimate factors that can have a substantial impact on building energy use. To overcome these limitations, this paper proposes a novel Data-driven Urban Energy Simulation (DUE-S) framework that integrates a network-based machine learning algorithm (ResNet) with engineering simulation to better understand how buildings consume energy on multiple temporal (hourly, daily, monthly) and spatial scales in a city (single building, block, urban). We validate the proposed DUE-S framework on a proof of concept case study of 22 densely located university buildings in California, USA. Our results indicate that the DUE-S framework is able to accurately predict urban scale energy consumption at hourly, daily and monthly intervals. Moreover, our results also demonstrate that the integration of data-driven and engineering simulation approaches can partially capture the inter-building energy dynamics and impacts of the urban context and merits future work to explore how they can be improved to predict sub-urban scale energy predictions (single building, block). In the end, successfully predicting and modeling the energy performance of urban buildings has the potential to inform the decision-making of a wide variety of urban sustainability stakeholders including architects, engineers and policymakers.

关键词： 反向传播人工神经网络   随机森林   支持向量机   K折交叉验证   混淆矩阵  

摘要： 吉林地理标志大米因其品质优,口感好,营养价值高,在市场上享有较高的声誉,研究地理标志大米产地确证技术具有重要意义。以吉林省柳河县和辉南县为研究区域,分别采集柳河地理标志大米样本62个,辉南地理标志大米样本58个,共120个样本。通过检测大米样本中矿物质元素[铜(Cu)、锌(Zn)、铁(Fe)、锰(Mn)、钾(K)、钙(Ca)、钠(Na)、镁(Mg)、铅(Pb)、镉(Cd)],利用反向传播人工神经网络(BackPropagation Artificial Neural Network,BP-ANN)、随机森林(Random Forest,RF)和支持向量机(Support Vector Machines,SVM)三种机器学习方法建立确证模型,并通过F评分(F-score)方法对矿物质元素进行特征提取,采用10次10折交叉验证和混淆矩阵对研究区域建立的产地确证模型进行评估比较。结果表明:单个Cu元素可作为代表该地区空间特征的典型变量。三种机器学习方法建立的产地确证模型均达到了良好的预测性能,其中BP-ANN方法使用Cu元素和Zn元素建立的分类模型准确率为99. 7%; SVM方法使用Cu元素、Zn元素和Pb元素建立的分类模型准确率为100%; RF方法使用Cu、Zn、Pb、Ca、Cd、K 6种元素建立的分类准确率为100%。RF模型和SVM模型整体分类效果优于BP-ANN模型,模型更稳定,更适合建立研究区域的产地确证模型。

关键词： 文本分类   中文分词   统计语言模型   神经网络模型  

摘要： 文本分类在文本信息处理过程中有关键作用,能帮助人们提取有用信息,有效组织文本信息。本文主要介绍了文本分类的背景、国内外的研究现状以及利用机器学习方法解决文本分类问题的一般步骤。文章第二部分对中文分词、特征向量提取、分类器训练和评估原理做了重点介绍,包括了自然语言处理的统计语言模型,机器学习的KNN、SVM、神经网络算法。

关键词： 统计学习   在线学习   再生核空间   随机动力学系统   鲁棒最优控制   模型预测控制  

摘要： 基于再生核Hilbert空间(reproducing kernel Hilbert space,RKHS)的统计学习模型被广泛应用于函数逼近、图像处理、模式识别和回归分析等领域,并且也在非线性随机动力学系统的辨识问题中有着很好的表现.本文提出一个基于鲁棒最优控制的RKHS模型学习方法,来实现对非线性随机动力学系统的高效在线建模.利用本文得到的关于再生核空间的一些理论结果,本文将随机动力学系统的在线学习问题转化为一组具有有界随机扰动的离散时变线性系统的输出反馈镇定问题,并利用模型预测控制技术来设计相应的控制算法和学习算法.与现有的RKHS模型学习方法相比,在不引入任何数据窗口原理、剪枝技术、学习步长的调整机制以及对噪声统计性质的假设的情形下,新方法可以在保证模型参数快速且鲁棒收敛的同时,实现对动力学系统的自适应高精度建模.此外,本文首次从最优控制的视角出发,研究动力学系统的在线核学习问题.在本文提出的研究框架下,现有各种控制技术可以被利用起来开发新的鲁棒学习方法,这也为核学习理论的研究和算法的开发提供一些新的思路.本文亦给出了数值算例和对比结果,用来说明新方法的有效性.

关键词： 疾病   机器学习   人工智能   临床决策支持  

摘要： 随着组学技术的不断发展,对于不同层次和类型的生物数据的获取方法日益成熟。在疾病诊治过程中会产生大量数据,通过机器学习等人工智能方法解析复杂、多维、多尺度的疾病大数据,构建临床决策支持工具,辅助医生寻找快速且有效的疾病诊疗方案是非常必要的。在此过程中,机器学习等人工智能方法的选择显得尤为重要。基于此,本文首先从类型和算法角度对临床决策支持领域中常用的机器学习等方法进行简要综述,分别介绍了支持向量机、逻辑回归、聚类算法、Bagging、随机森林和深度学习,对机器学习等方法在临床决策支持中的应用做了相应总结和分类,并对它们的优势和不足分别进行讨论和阐述,为临床决策支持中机器学习等人工智能方法的选择提供有效参考。

关键词： 脑电波   行为决策   熵   机器学习方法  

摘要： 随着年龄的增长大脑的功能会衰退.可以通过决策实验获取年轻人和中老年人的行为决策脑电信号,定量分析大脑随年龄增长而出现的变化.本文提出了一种基于熵的脑电波刻画方法,结合机器学习的方法能够比较准确地诊断人类的大脑年龄.研究结果表明,年轻人和中老年人在原始脑电信号上存在明显的电压差异.此外,XGBoost的分类性能优于Logistic Regression(LR)方法的,最高精度可达91%,比LR的高出7.22%.通过XGBoost增益指标对特征重要性排序,发现在决策过程中受试者大脑中央和颞区域的决策反应差异最大,而机器学习分类方法更容易在这些特征区域做出更好的分类.