基于多目标多样性回声状态网络的时间序列分析

发布时间：2018-05-06 09:07

本文选题：时间序列预测 + 回声状态网络　；参考：《中国科学技术大学》2017年硕士论文

【摘要】：时间序列数据广泛存在于电子商务,金融,视频活动分析等任务中,针对时序数据的分析和预测是一项十分重要且有挑战性的工作。对时间序列分析的难点有二,其一是:时间序列数据具有时序性。其二是:时间序列数据通常包含较多的噪声。回声状态网络(Echo State Network,ESN)是目前一种流行的时间序列分析模型,它的储蓄池(相当于传统神经网络的隐藏层)的循环连接使其具有一定的"记忆"能力,可以很好地拟合出时间序列数据的时序关系。由于回声状态网络储蓄池中的节点个数较多,当它被用来拟合噪声较多的数据时,容易造成过拟合,影响模型的预测能力。传统ESN模型中,ESN输入层到储蓄池、储蓄池内部节点之间连接权重随机生成,ESN的性能依赖于这种随机性,导致它具有不稳定性。为了得到合适的ESN,通常需要不断地随机生成ESN,直到产生合适模型为止,传统方法无法保证新生成的ESN优于之前的ESN。基于回声状态网络的上述缺点,本文提出了多目标多样性回声状态网络(Multi-objective Diversified Echo State Network,MODESN),MODESN 定义了 ESN 多样性。ESN 多样性通过考虑储蓄池中节点之间的冗余度,从而尽可能地避免过拟合情况发生。MODESN利用多目标遗传算法同时对ESN多样性和预测准确率进行优化,使得新生成的ESN模型向期望的方向演化,从而避免了传统生成ESN方法的随机性。本文的主要工作可以总结如下:(1)本文定义了 ESN多样性,通过优化ESN多样来优化ESN结构,从而尽可能避免过拟合。(2)本文改进了一种多目标遗传算法,并用其对ESN多样性和预测准确率同时进行优化,规范了 ESN演化方向。(3)本文将日本蜡烛图技术与MODESN模型结合,降低了特征数量,避免引入不必要的噪声。
[Abstract]:Time series data are widely used in e-commerce, finance, video activity analysis and other tasks. The analysis and prediction of time series data is a very important and challenging task. There are two difficulties in time series analysis. One is that the time series data are temporal. The second is that time series data usually contain more noise. Echo State Network (ESN) is a popular time series analysis model. Its storage pool (equivalent to the hidden layer of the traditional neural network) is cyclically connected to enable it to have a certain "memory" capability. The temporal relationship of time series data can be fitted well. Because of the large number of nodes in the echo state network savings pool, when it is used to fit the noisy data, it is easy to cause over-fitting, which affects the prediction ability of the model. In the traditional ESN model, the performance of randomly generating the ESN input layer from the input layer to the savings pool is dependent on this randomness, which leads to its instability. In order to obtain a suitable ESNs, it is usually necessary to generate them at random until the appropriate models are generated. The traditional method can not guarantee that the newly generated ESN is superior to the previous ones. Based on the above disadvantages of echo state network, this paper proposes a multi-objective Diversified Echo State network named Multi-objective Diversified Echo State Network / MODESN which defines ESN diversity by considering the redundancy between nodes in the storage pool. In order to avoid overfitting as far as possible. MODESN optimizes the diversity and prediction accuracy of ESN using multi-objective genetic algorithm at the same time, so that the newly generated ESN model evolves in the desired direction. Thus, the randomness of the traditional ESN generation method is avoided. The main work of this paper can be summarized as follows: 1) this paper defines the diversity of ESN, optimizes the structure of ESN by optimizing the diversity of ESN, and avoids overfitting. 2) this paper improves a multi-objective genetic algorithm. In this paper, we combine the Japanese candle chart technology with the MODESN model to reduce the number of features and avoid the introduction of unnecessary noise by optimizing the diversity and prediction accuracy of ESN at the same time, and standardizing the evolution direction of ESN.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：O211.61

【参考文献】