北京市生态安全格局保护紧迫性分级

以北京市为例, 通过粒度反推法和景观连通性识别重要生态用地, 通过邻域分析法构建生态阻力面, 采用MCR模型构建北京生态安全格局, 利用人工神经网络计算建设用地扩张概率, 对构建的生态安全格局进行保护紧迫性分级。结果表明, 北京市重要生态用地面积为 6488.53 km², 主要分布在西部和东北部山区, 小部分位于东南平原。一级生态用地面积为4482.48 km², 主要分布在西北和西南地区; 二级生态用地面积为1338.27 km², 主要分布在西部远郊山区、西南部近郊平原和东南部平原; 三级生态用地面积为669.77 km², 主要分布在中北部和东南部平原。重要生态廊道长度为2410.47 km, 分布在远郊城镇以及城乡交界处。一级生态廊道长度为1477.63 km, 主要分布在西北和西南远郊山区; 二级生态廊道长度为390.91 km, 主要分布在远郊平原的生态用地附近; 三级生态廊道长度为 541.93 km, 主要分布在城市中心的周边地区, 呈环状包围中心城区。     

Identifying US business cycle regimes using dynamic factors and neural network models

We use dynamic factors and neural network models to identify current and past states (instead of future) of the US business cycle. In the first step, we reduce noise in data by using a moving average filter. Dynamic factors are then extracted from a large-scale data set consisted of more than 100 variables. In the last step, these dynamic factors are fed into the neural network model for predicting business cycle regimes. We show that our proposed method follows US business cycle regimes quite accurately in-sample and out-of-sample without taking account of the historical data availability. Our results also indicate that noise reduction is an important step for business cycle prediction. Furthermore, using pseudo real time and vintage data, we show that our neural network model identifies turning points quite accurately and very quickly in real time.     

Neural network structure identification in inflation forecasting

Neural networks (NNs) are appropriate to use in time series analysis under conditions of unfulfilled assumptions, i.e., non‐normality and nonlinearity. The aim of this paper is to propose means of addressing identified shortcomings with the objective of identifying the NN structure for inflation forecasting. The research is based on a theoretical model that includes the characteristics of demand‐pull and cost‐push inflation; i.e., it uses the labor market, financial and external factors, and lagged inflation variables. It is conducted at the aggregate level of euro area countries from January 1999 to January 2017. Based on the estimated 90 feedforward NNs (FNNs) and 450 Jordan NNs (JNNs), which differ in variable parameters (number of iterations, learning rate, initial weight value intervals, number of hidden neurons, and weight value of the context unit), the mean square error (MSE), and the Akaike Information Criterion (AIC) are calculated for two periods: in‐the‐sample and out‐of‐sample. Ranking NNs simultaneously on both periods according to either MSE or AIC does not lead to the selection of the ‘best’ NN because the optimal NN in‐the‐sample, based on MSE and/or AIC criteria, often has high out‐of‐sample values of both indicators. To achieve the best compromise solution, i.e., to select an optimal NN, the preference ranking organization method for enrichment of evaluations (PROMETHEE) is used. Comparing the optimal FNN and JNN, i.e., FNN(4,5,1) and JNN(4,3,1), it is concluded that under approximately equal conditions, fewer hidden layer neurons are required in JNN than in FNN, confirming that JNN is parsimonious compared to FNN. Moreover, JNN has a better forecasting performance than FNN.     

小样本条件下SCGAN+CNN低分辨雷达目标一步识别算法

现有低分辨雷达目标识别方法，通常采用先特征提取、再进行目标分类的两步识别算法，这种算法存在识别率难以提高和方法泛化性不足的问题，对此，提出一种增强条件生成对抗网络（strengthening condition generative adversarial network，SCGAN）+卷积神经网络（convolutional neural network，CNN）的低分辨雷达目标一步识别算法。该算法利用CNN自动获取采样数据深层本质特征，无需特征提取，实现对目标的一步识别。为进一步提高小样本条件下的识别效果，基于CGAN理论来提高样本在特征空间的覆盖程度，并对CGAN的判别器进行改进，在损失函数中增加混叠惩戒项，通过SCGAN生成不混叠的生成样本来更好地训练CNN，提高其在小样本条件下的识别能力。仿真对比实验校验了一步识别算法较传统两步识别算法的优越性，以及SCGAN+CNN的低分辨雷达目标一步识别算法在小样本条件下的有效性。     

Steganalysis of neural networks based on parameter statistical bias

Many pretrained deep learning models have been released to help engineers and researchers develop deep learning-based systems or conduct research with minimall effort. Previous work has shown that at secret message can be embedded in neural network parameters without compromising the accuracy of the model. Malicious developers can, therefore, hide malware or other baneful information in pretrained models, causing harm to society. Hence, reliable detection of these vicious pretrained models is urgently needed. We analyze existing approaches for hiding messages and find that they will ineluctably cause biases in the parameter statistics. Therefore, we propose steganalysis methods for steganography on neural network parameters that extract statistics from benign and malicious models and build classifiers based on the extracted statistics. To the best of our knowledge, this is the first study on neural network steganalysis. The experimental results reveal that our proposed algorithm can effectively detect a model with an embedded message. Notably, our detection methods are still valid in cases where the payload of the stego model is low.     

混合动力汽车深度强化学习分层能量管理策略

为了提高混合动力汽车的燃油经济性和控制策略的稳定性,以第三代普锐斯混联式混合动力汽车作为研究对象,提出了一种等效燃油消耗最小策略(equivalent fuel consumption minimization strategy,ECMS)与深度强化学习方法(deep feinforcement learning,DRL)结合的分层能量管理策略。仿真结果证明,该分层控制策略不仅可以让强化学习中的智能体在无模型的情况下实现自适应节能控制,而且能保证混合动力汽车在所有工况下的SOC都满足约束限制。与基于规则的能量管理策略相比,此分层控制策略可以将燃油经济性提高20.83%~32.66%;增加智能体对车速的预测信息,可进一步降低5.12%的燃油消耗;与没有分层的深度强化学习策略相比,此策略可将燃油经济性提高8.04%;与使用SOC偏移惩罚的自适应等效燃油消耗最小策略(A-ECMS)相比,此策略下的燃油经济性将提高5.81%~16.18%。     

基于小波分析的神经网络识别毒剂的研究

为快速、准确地识别毒剂,在分析神经网络识别毒剂基本方法的基础上,建立了带有偏差单元的递归神经网络识别毒剂模型,包括神经网络识别毒剂的学习算法和基于小波分析的毒剂特征提取. 通过剖析神经网络识别毒剂模型,设计了神经网络识别毒剂的软件,实现了神经网络对毒剂的识别. 用沙林模拟数据进行了测试和分析,结果表明,利用与化学传感器相联结的神经网络识别毒剂,是实现毒剂识别自动化、智能化的一种有效方法.     

电磁带隙结构的优化设计

运用人工神经网络对电磁带隙结构进行优化设计,并计算了几组检验样本对神经网络的精确度进行检验。训练样本和检验样本的获得均使用了基于有限元方法的HFSS商业软件,所表述的方法具有缩短设计周期,降低研发成本的优点。     

人工神经网络预测混凝土柱屈服性能

建立了一种基于人工神经网络的矩形混凝土柱屈服性能预测方法.该方法采用经验模型进行柱屈服性能影响因素的分析来确定该神经网络的输入参数,并通过敏感性分析验证了所选神经网络输入参数的合理性.为验证该方法的可行性与有效性,通过对PEER 210组矩形混凝土柱的屈服性能进行预测分析并与经验预测模型的预测结果进行比较.比较分析结果表明:神经网络预测结果与实验结果吻合程度远高于其他经验预测模型;同时也证实该方法在实验数据稀少的情况下为预测结构在地震作用下的性能提供一条新途径.     

Software Maintainability Prediction with UML Class Diagram

Software system can be classified into many function modules from the perspective of user. Unified modeling language( UML) class diagram of each function module was extracted,and design characteristic metrics which influenced software maintainability were selected based on UML class diagram.Choosing metrics of UML class diagram as predictors,and mean maintenance time of function module was regarded as software maintainability parameter. Software maintainability models were built by using back propagation( BP) neural network and radial basis function( RBF) neural network, respectively and were simulated by MATLAB. In order to evaluate the performance of models,the training results were analyzed and compared with leaveone-out cross-validation and model performance evaluation criterion. The result indicated that RBF arithmetic was superior to BP arithmetic in predicting software maintainability.