基于深度学习的盲文识别方法

提出一种基于深度学习的盲文点字识别方法,利用深度模型--堆叠去噪自动编码器（Stack Denoising AutoEncoder,SDAE）解决盲文识别中特征的自动提取与降维等问题。在构建深度模型过程中,采用非监督贪婪逐层训练算法（Greedy Layer Wise Unsupervised Learning Algorithm）初始化网络权重,使用反向传播算法优化网络参数。利用SDAE自动学习盲文点字图片特征,使用Softmax分类器进行识别。实验结果表明,本文所提方法较之传统方法,可以有效解决样本特征的自动学习与特征降维等问题,操作更为简易,并能获得满意的识别结果。     

弱监督分层深度学习的车辆识别算法

针对已有分类器在结构形式和训练方法的不足,构建了一个以二维深度置信网络（2D deep belief networks,2D DBN）为架构的弱监督分层深度学习车辆识别算法。首先,将传统一维的深度置信网络(Deep belief networks,DBN)扩展成2D-DBN,并构建相应分类器结构,从而能够直接以二维图像像素矩阵作为输入; 其次,在传统无监督训练的目标函数中,引入了一个具有适当权重的判别度正则化项,将原有无监督训练转化为带有较弱监督性的弱监督训练方式,从而使提取的特征较传统无监督特征更具判别性。多组对比实验表明,本文所提算法在识别率等指标上要优于已有深度学习算法。     

改进DDPG算法在自动驾驶中的应用

深度确定性策略梯度算法（Deep Deterministic Policy Gradient，DDPG）作为深度强化学习中的经典算法，在连续控制问题上有着较大的优势，被应用于自动驾驶领域。针对DDPG缺少策略动作过滤导致的非法策略比例较高引起的训练效率低、收敛速度慢等问题，提出基于失败经验纠错的深度确定性策略梯度算法。通过分离经验缓存池，根据驾驶表现选择失败数据训练，并将策略网络单输出转化为油门和刹车控制量，通过正态分布噪声改善探索策略。TORCS平台仿真实验表明，所提算法相对于DDPG算法与DQN（Deep Q-learning Network）算法，训练效率明显提升，非法驾驶策略降低为0。     

集成的深度强化学习投资组合模型

投资组合问题是量化交易领域中的热点问题。针对现有基于深度强化学习的投资组合模型无法实现自适应的交易策略和有效利用有监督信息的缺陷,提出一种集成的深度强化学习投资组合模型（IDRLPM）。首先,采用多智能体方法构造多个基智能体并设计不同交易风格的奖励函数,以表示不同的交易策略;其次,利用集成学习方法对基智能体的策略网络进行特征融合,得到自适应市场环境的集成智能体;然后,在集成智能体中嵌入基于卷积块注意力模块（CBAM）的趋势预测网络,趋势预测网络输出引导集成策略网络自适应选择交易比重;最后,在有监督深度学习和强化学习交替迭代训练下,IDRLPM有效利用训练数据中的监督信息以增强模型盈利能力。在上证50的成分股和中证500的成分股数据集中,IDRLPM的夏普比率（SR）达到了1.87和1.88,累计收益（CR）达到了2.02和1.34;相较于集合式的深度强化学习（EDRL）交易模型,SR提高了105%和55%,CR提高了124%和79%。实验结果表明,IDRLPM能够有效解决投资组合问题。     

引入通信与探索的多智能体强化学习QMIX算法

非平稳性问题是多智能体环境中深度学习面临的主要挑战之一,它打破了大多数单智能体强化学习算法都遵循的马尔可夫假设,使每个智能体在学习过程中都有可能会陷入由其他智能体所创建的环境而导致无终止的循环。为解决上述问题,研究了中心式训练分布式执行（CTDE）架构在强化学习中的实现方法,并分别从智能体间通信和智能体探索这两个角度入手,采用通过方差控制的强化学习算法（VBC）并引入好奇心机制来改进QMIX算法。通过星际争霸Ⅱ学习环境（SC2LE）中的微操场景对所提算法加以验证。实验结果表明,与QMIX算法相比,所提算法的性能有所提升,并且能够得到收敛速度更快的训练模型。     

基于加权密集连接卷积网络的深度强化学习方法

针对深度强化学习中卷积神经网络（CNN）层数过深导致的梯度消失问题，提出一种将密集连接卷积网络应用于强化学习的方法。首先，利用密集连接卷积网络中的跨层连接结构进行图像特征的有效提取；然后，在密集连接卷积网络中加入权重系数，加权密集连接卷积网络中的每一层都接收到前面几层产生的所有特征图，且之前所有层在跨层连接中被赋予不同的初始权重；最后，在训练中动态调整每层的权重，从而更加有效地提取特征。与常规深度强化学习方法相比，在GridWorld仿真实验中，在相同训练步数内的平均奖励值提升了85.67%；在FlappyBird仿真中，平均奖励值提升了55.05%。实验结果表明所提方法能在不同难度的游戏仿真实验中获得更好的性能。     

深度森林在用户购买预测中的应用

针对电商大数据时代用户未来购买行为预测，在京东平台真实数据集上，提出时间滑动窗口技术和窗口权重递减设置，从五方面构建整体用户行为特征，综合考虑深度学习的表征学习能力和集成学习的训练效率，引入多层异源集成算法，将随机森林、XGBoost等多种算法进行组合，搭建基于深度森林模型的用户购买行为预测算法框架，实现准确高效的用户购买预测结果。算法训练时间为68 s,预测准确率达89.3%,相对于集成学习算法和深度神经网络模型取得了更好的效果。     

基于双重注意力机制的异步优势行动者评论家算法

深度强化学习是目前机器学习领域发展最快的技术之一.传统的深度强化学习方法在处理高维度大状态的空间任务时,庞大的计算量导致其训练时间过长.虽然异步深度强化学习利用异步方法极大缩短了训练时间,但会忽略某些更具价值的图像区域和图像特征.针对上述问题,本文提出了一种基于双重注意力机制的异步优势行动者评论家算法.新算法利用特征注意力机制和视觉注意力机制来改进传统的异步深度强化学习模型.其中,特征注意力机制为卷积神经网络卷积后的所有特征图设置不同的权重,使得智能体聚焦于重要的图像特征;同时,视觉注意力机制为图像不同区域设置权重参数,权重高的区域表示该区域信息对智能体后续的策略学习有重要价值,帮助智能体更高效地学习到最优策略.新算法引入双重注意力机制,从表层和深层两个角度对图像进行编码表征,帮助智能体将聚焦点集中在重要的图像区域和图像特征上.最后,通过Atari 2600部分经典实验验证了基于双重注意力机制的异步优势行动者评论家算法的有效性.     

基于深度强化学习的机械臂控制快速训练方法

人工智能在机器人控制中得到广泛应用,机器人控制算法也逐渐从模型驱动转变为数据驱动。深度强化学习算法可在复杂环境中感知并决策,能够解决高维度和连续状态空间下的机械臂控制问题。然而,目前深度强化学习中数据驱动的训练过程非常依赖计算机GPU算力,且训练时间成本较大。提出基于深度强化学习的先简化模型（2D模型）再复杂模型（3D模型）的机械臂控制快速训练方法。采用深度确定性策略梯度算法代替机械臂传统控制算法中的逆运动学解算方法,直接通过数据驱动的训练过程控制机械臂末端到达目标位置,从而减小训练时间成本。同时,对于状态向量和奖励函数形式,使用不同的设置方式。将最终训练得到的算法模型在真实机械臂上进行实现和验证,结果表明,其控制效果达到了分拣物品的应用要求,相比于直接在3D模型中的训练,能够缩短近52%的平均训练时长。     

融合知识图谱与Bert+CNN的图书文本分类研究

基于海量图书的分类需求，提出了一种融合知识图谱（Knowledge Graph, KG）与转换器双向编码器（Bidirectional Encoder Representation from Transformers, Bert）的图书文本分类模型。通过构建面向图书领域知识图谱扩展图书文本的语义信息，并且使用深度学习的方法获取文本深层语义信息，将扩展语义信息与深层语义信息相结合后，通过TextCNN进行图书分类。经研究试验表明，融合知识图谱与深度学习的图书文本分类算法相对于只使用深度学习进行图书分类算法，前者的分类效果更佳。     

基于PSO的考虑完整费用的证券组合优化研究

通过分析中国证券市场证券交易不可拆分、不能卖空的特点以及现存的各种交易费用,建立一个考虑完整交易费用的证券投资组合优化模型,同时给出一个应用粒子群算法(PSO)求解的实例。结果证明该证券投资组合优化模型的完整性和有效性,也表明PSO算法可以快速准确地求解证券投资组合优化问题。     

大数据分析的应用案例——投资模型的稳健性

交易模型的稳健性,指的是该模型的利润率曲线的波动性较小,没有大起大落。针对一个基于支持向量回归（SVR）技术的算法交易模型的稳健性问题,提出了使用若干导出指标训练统一的交易模型的策略,以及投资组合多样化的方法。首先,介绍基于支持向量回归技术的算法交易模型;然后,基于常用指标,构造了若干导出指标,用于股票价格的短期预测。这些指标,刻画了近期价格运动的典型模式、超买/超卖市场状态,以及背离市场状态。对这些指标进行了规范化,用于训练交易模型,使得模型可以泛化到不同的股票;最后,设计了投资组合多样化方法。在投资组合里,各个股票之间的相关性,有时会导致较大的投资损失;因为具有较强相关关系的股票,其价格朝相同方向变化。如果交易模型预测的价格走势不正确,引起止损操作,那么这些具有较强相关关系的股票,将引发雪崩式的止损,于是导致损失加剧。把股票根据相似性聚类到不同类别,通过从不同聚类类别中选择若干股票来构成多样化的投资组合,其中,股票的相似性,通过交易模型在不同股票上近期的利润曲线的相似度进行计算。在900只股票10年的价格大数据上进行了实验,实验结果显示,交易模型能够获得超过定期存款的超额利润率,年化利润率为8.06%。交易模型的最大回撤由13.23%降为5.32%,夏普指数由81.23%提高到88.79%,交易模型的利润率曲线波动性降低,说明交易模型的稳健性获得了提高。     

A hybrid stock selection model using genetic algorithms and support vector regression

In the areas of investment research and applications, feasible quantitative models include methodologies stemming from soft computing for prediction of financial time series, multi-objective optimization of investment return and risk reduction, as well as selection of investment instruments for portfolio management based on asset ranking using a variety of input variables and historical data, etc. Among all these, stock selection has long been identified as a challenging and important task. This line of research is highly contingent upon reliable stock ranking for successful portfolio construction. Recent advances in machine learning and data mining are leading to significant opportunities to solve these problems more effectively. In this study, we aim at developing a methodology for effective stock selection using support vector regression (SVR) as well as genetic algorithms (GAs). We first employ the SVR method to generate surrogates for actual stock returns that in turn serve to provide reliable rankings of stocks. Top-ranked stocks can thus be selected to form a portfolio. On top of this model, the GA is employed for the optimization of model parameters, and feature selection to acquire optimal subsets of input variables to the SVR model. We will show that the investment returns provided by our proposed methodology significantly outperform the benchmark. Based upon these promising results, we expect this hybrid GA-SVR methodology to advance the research in soft computing for finance and provide an effective solution to stock selection in practice.     

A Bayesian information criterion for portfolio selection

The mean-variance theory of Markowitz (1952) indicates that large investment portfolios naturally provide better risk diversification than small ones. However, due to parameter estimation errors, one may find ambiguous results in practice. Hence, it is essential to identify relevant stocks to alleviate the impact of estimation error in portfolio selection. To this end, we propose a linkage condition to link the relevant and irrelevant stock returns via their conditional regression relationship. Subsequently, we obtain a BIC selection criterion that enables us to identify relevant stocks consistently. Numerical studies indicate that BIC outperforms commonly used portfolio strategies in the literature.     

基于压缩感知和深度小波网络的列车故障识别

针对列车走行部故障振动数据无监督特征学习的难点，提出了一种基于压缩感知和深度小波神经网络（CS-DWNN）的列车故障识别方法。首先，对采集得到的列车走行部振动信号利用高斯随机矩阵进行压缩采样；其次，构建以改进小波自编码器（WAE）为基础的深层小波网络，将压缩后的信号直接输入网络进行自动逐层特征提取；最后，用DWNN学习到的多层特征分别训练多个深度支持向量机（DSVM）和深度森林（DF）分类器，并将识别结果进行集成。该方法利用深层小波网络从压缩信号中自动挖掘隐藏的故障信息，受先验知识和主观影响较小，并且避免了复杂的人工特征提取过程。实验结果表明，CS-DWNN方法取得了99.16%的平均诊断正确率，能够有效识别列车走行部的3种常见故障，识别能力优于传统的人工神经网络（ANN）、支持向量机（SVM）等方法和深度信念网络（DBN）、堆栈降噪自编码器（SDAE）等深度学习模型。     

An integrated portfolio optimisation procedure based on data envelopment analysis,artificial bee colony algorithm and genetic programming

Portfolio optimisation is an important issue in the field of investment/financial decision-making and has received considerable attention from both researchers and practitioners. However, besides portfolio optimisation, a complete investment procedure should also include the selection of profitable investment targets and determine the optimal timing for buying/selling the investment targets. In this study, an integrated procedure using data envelopment analysis (DEA), artificial bee colony (ABC) and genetic programming (GP) is proposed to resolve a portfolio optimisation problem. The proposed procedure is evaluated through a case study on investing in stocks in the semiconductor sub-section of the Taiwan stock market for 4 years. The potential average 6-month return on investment of 9.31% from 1 November 2007 to 31 October 2011 indicates that the proposed procedure can be considered a feasible and effective tool for making outstanding investment plans, and thus making profits in the Taiwan stock market. Moreover, it is a strategy that can help investors to make profits even when the overall stock market suffers a loss.     

Score-hesitation trade-off and portfolio selection under intuitionistic fuzzy environment

Qualitative portfolio selection approach is a suitable technique for obtaining an optimal portfolio when quantitative data are unavailable and traditional portfolio models are ineffective. However, few studies focus on this issue. This study addresses the lack of research by defining the score-hesitation trade-off rule and introducing the intuitionistic fuzzy set (IFS), based on which an intuitionistic fuzzy portfolio selection (IFPS) model is proposed. The IFS is introduced because of its comprehensive consideration of preference and nonpreference, and is used to represent qualitatively evaluated information from investors and experts. Furthermore, an intuitionistic fuzzy investment scenario is established and a trisection approach is designed to distinguish three types of risk investors, based on which three corresponding IFPS models are constructed. After this, a portfolio selection process under the intuitionistic fuzzy environment is provided, and a simple example is given to show the application of the process. In addition, the investment opportunities and efficient frontier of the IFPS model are investigated to demonstrate the effectiveness of the proposed portfolio selection model. Finally, an example of calculating optimal investment ratios and selecting an optimal portfolio for four newly listed stocks in China is provided to demonstrate the feasibility and practicability of the proposed approaches.     

基于时序超图卷积神经网络的股票趋势预测方法

传统的股票预测方法大多基于时间序列模型,忽视了股票之间复杂的关系,并且该关系往往超出成对连接,例如同行业板块内股票或者基金持仓多支股票.针对该问题,提出一种基于时序超图卷积神经网络(HGCN)的股价走势预测方法,根据金融投资事实构造超图模型以拟合股票之间的多元关系,该模型包括两大组件:门控循环单元(GRU)网络和超图卷...     

Application of VPRS model with enhanced threshold parameter selection mechanism to automatic stock market forecasting and portfolio selection

This study proposes a technique based upon Fuzzy C-Means (FCM) classification theory and related fuzzy theories for choosing an appropriate value of the Variable Precision Rough Set (VPRS) threshold parameter (β) when applied to the classification of continuous information systems. The VPRS model is then combined with a moving Average Autoregressive Exogenous (ARX) prediction model and Grey Systems theory to create an automatic stock market forecasting and portfolio selection mechanism. In the proposed mechanism, financial data are collected automatically every quarter and are input to an ARX prediction model to forecast the future trends of the collected data over the next quarter or half-year period. The forecast data are then reduced using a GM(1, N) model, classified using a FCM clustering algorithm, and then supplied to a VPRS classification module which selects appropriate investment stocks in accordance with a pre-determined set of decision-making rules. Finally, a grey relational analysis technique is employed to weight the selected stocks in such a way as to maximize the rate of return of the stock portfolio. The validity of the proposed approach is demonstrated using electronic stock data extracted from the financial database maintained by the Taiwan Economic Journal (TEJ). The portfolio results obtained using the proposed hybrid model are compared with those obtained using a Rough Set (RS) selection model. The effects of the number of attributes of the RS lower approximation set and VPRS β-lower approximation set on the classification are systematically examined and compared. Overall, the results show that the proposed stock forecasting and stock selection mechanism not only yields a greater number of selected stocks in the β-lower approximation set than in the RS approximation set, but also yields a greater rate of return.     

A stock market portfolio recommender system based on association rule mining

We propose a stock market portfolio recommender system based on association rule mining (ARM) that analyzes stock data and suggests a ranked basket of stocks. The objective of this recommender system is to support stock market traders, individual investors and fund managers in their decisions by suggesting investment in a group of equity stocks when strong evidence of possible profit from these transactions is available.Our system is different compared to existing systems because it finds the correlation between stocks and recommends a portfolio. Existing techniques recommend buying or selling a single stock and do not recommend a portfolio.We have used the support confidence framework for generating association rules. The use of traditional ARM is infeasible because the number of association rules is exponential and finding relevant rules from this set is difficult. Therefore ARM techniques have been augmented with domain specific techniques like formation of thematical sectors, use of cross-sector and intra-sector rules to overcome the disadvantages of traditional ARM.We have implemented novel methods like using fuzzy logic and the concept of time lags to generate datasets from actual data of stock prices.Thorough experimentation has been performed on a variety of datasets like the BSE-30 sensitive Index, the S&P CNX Nifty or NSE-50, S&P CNX-100 and DOW-30 Industrial Average. We have compared the returns of our recommender system with the returns obtained from the top-5 mutual funds in India. The results of our system have surpassed the results from the mutual funds for all the datasets.Our approach demonstrates the application of soft computing techniques like ARM and fuzzy classification in the design of recommender systems.