区域物流需求预测现状和发展研究

  • 格式:doc
  • 大小:33.50 KB
  • 文档页数:4

从宏观产业经济发展的角度看,各种区域物流发展政策、区域物流规划的制定等都离不开区域物流需求的定量分析,在规划中遇到的首要问题就是对宏观物流市场规模的预测。

区域物流需求预测可为国家和地区物流经济主管部门制定未来物流发展的科学发展战略规划以及切实可行的市场开拓策略提供依据,近而为宏观产业经济政策的制定提供参考。

另外,政府可以通过区域物流需求预测来评估物流行业对当地经济发展的总体贡献,从而制定物流行业发展政策,并引导物流市场资源的合理利用与优化配置。

从微观角度看,物流企业需根据对物流需求预测合理地配置有限的资源,以期最大限度降低投资风险和获得最大收益。

国内外研究现状区域物流需求预测研究开始于上世纪90年代。

我国学者在本世纪初开始研究区域物流需求预测方法。

经过近20年的发展,区域物流需求预测研究取得了很大进步。

通过对目前区域物流需求预测方法研究发现,目前主要把区域物流需求预测看作回归问题。

根据发展历程和智能化程度的高低,大致可以划分为以下三个阶段:第一阶段主要采用基于传统统计学的预测方法,这是本世纪初区域物流需求预测主要采用的方法。

该类主要方法包括:投入产出模型、回归分析法、货运强度法、弹性系数法、聚类法、灰色理论模型、马尔可夫链、时空多项概率模型和决策支持系统等。

该类方法的主要特点可以对定序和线性的数据进行处理,且对于构造的模型有较强的解释性。

随着区域物流需求预测研究的深入,大多数研究方法的不足就逐渐暴露出来,主要体现在以下三个方面:第一,真实的区域物流需求数据样本非常少且难以收集,这极大地影响了预测方法的验证效果。

第二,在处理高维度、含有非线性关系、程非正态分布、有时间顺序的区域物流需求数据时,其效果不理想。

第三,不能保证学习和泛化能力,缺乏灵活性,而且,整个处理过程按规定的步骤进行,对所有数据一视同仁地进行同样的处理而不管是否需要进行这些处理。

后面两个问题的产生促使了人们考虑在物流需求预测中引人人工智能技术,以改善预测模型的性能和提高预测准确率。

目前学者采用的人工智能方法主要是人工神经网络(ANN)及其改进型。

人工智能预测方法的引入使传统预测方法中融人了人的智能因素,如神经网络的学习和泛化能力,专家系统的推理规则等,这样区域物流需求预测技术又向前迈进了一大步。

但是,从前面的分析可以看出:第一,不能从理论上保证预测模型的泛化能力,这使得对于经过训练后的预测模型,对于新的物流需求量数据集没有稳定的预测效果。

第二,在学习样本数量有限时,学习过程误差易收敛于局部极小点,学习精度难以保证;学习样本变量很多时,又陷入“维数灾难”。

第三,主要依靠的是经验风险最小化原则,其重要的缺点为:利用经验风险代替期望风险来选择决策函数,并没有经过严格的证明与充分的论证,只是一种认为合理的方法;容易产生“过学习”问题,这是盲目追求小误差而导致泛化能力下降的必然结果。

因此,为了进一步地提高预测模型的精度和稳定性,近年来,研究人员开始探索基于结构风险最小化原则的预测方法,这使区域物流需求预测研究跨入到了第三阶段,即基于统计学习理论的预测方法。

该类方法主要以支持向量机(Support Vector Machine,SVM)为代表。

在引入基于统计学习理论的预测方法后,尽管区域物流需求研究者做了不少工作,取得了一些研究成果,但这些方法均存在这样或那样的不足,或者说在区域物流需求预测中还存在着一些迄待解决的问题,这些不足或问题主要集中在以下几个方面:第一,.在学习样本数量有限时,学习过程误差易收敛于局部极小点,学习精度难以保证;学习样本变量很多时,又陷入“维数灾难”。

第二,主要依靠的是经验风险最小化原则。

因此,不能从理论上保证预测模型的泛化能力,这使得对于经过训练后的预测模型,对于新的物流需求量数据集没有稳定的预测效果。

第三,由于区域物流概念在我国刚引进不久,很多物流标准没统一,缺乏直接的统计数据,预测研究所采用的历史数据,只能采用相关的货运量来代替,这影响了预测方法的实证研究。

特别是在国内,大多数文献只是对方法的综述,几乎没有利用实际的区域物流数据进行预测分析,这使得众多的研究缺乏实际意义。

第四,预测所依靠的区域物流指标的选择方法,主要依靠行业研究者和从业者经验、主成分分析、因子分析等传统方法,缺乏新的方法研究。

第五,这些方法中,人工智能方法和支持向量机方法虽然在预测精度方面具有一定优势,但是不能对模型进行有效的解释。

未来的研究方向通过上述分析可以看出,今后研究的目标应该是在建立统一的区域物流需求预测模型结构框架基础上,利用新的回归方法增加预测模型的预测精度和增强预测模型的可解释性。

区域物流需求预测模型结构。

我们认为,在区域物流需求预测模型中,支持向量机的决策函数中的可以描述输入空间数据的属性关系,其值大小能反映对区域物流需求预测效果影响程度。

在前期工作中,黄虎利用支持向量回归机(SVR)建立了区域物流需求预测模型,实证结果表明该方法能取得较高的预测精度,但模型的解释性较差。

因此,引入作为属性权重来增强预测模型的解释性值得进一步深入的研究。

基于集成理论的区域物流需求预测模型的研究。

传统区域物流需求预测模型只是由单个的预测方法构成,但如何找到由单个方法建立的优化模型却是一个NP问题(如神经网络的网络配置,支持向量机的核函数选择和决策树参数选择等)。

在实际应用中,由于缺乏问题的先验知识,往往很难找到理想的优化模型,而这影响了预测模型泛化能力的提高。

前人的研究工作表明,即使在集成原理不清楚的情况下,也可以通过对一组方法进行投票或平均,提高学习系统的处理能力。

因此,引入集成学习理论的方法,避开以参数和函数选择为基础的模型优化方式,是提高预测模型泛化能力的一条新途径。

基于损失函数的区域物流需求预测模型研究。

在使用模型进行预测的过程中,关于精度分析和精度评价的方法文献很多,如建立预测模型的最优准则“误差平方和最小”、“误差的绝对值之和最小”等。

但这些方法和准则大都只考虑了误差的大小却没有考虑误差的方向,对正、负误差所产生的不同结果也没有充分考虑。

在前期工作中,我们根据区域物流需求预测特点,对于某种预测方法M定义了一个总损失函数c=∑c(q),其中q(t=1,2,…,n)为预测误差序列,c(e)为损失函数,用以衡量预测误差e1对决策所造成损失的大小。

由于使预测方法M的平均损失达到最小与使c=∑c(ei)达到最小是等价的,因此得到判别准则:对于给定的损失函数c(e;),在可供选择的预测方法中,使总损失c=∑c(q)取得最小值的方法是最优的预测方法。

当然,前期研究对于损失函数没有给出一个统一的表达式,且没有在实际的环境中进行验证,因此这部分内容还值得进一步深入的研究。

Industrial economy development from the macroscopic point of view, all sorts of regional logisticsdevelopment policy, the establishment of regional logistics planning is inseparable from the regional logistics demand of quantitative analysis, the primary problem is encountered in the planning of macroscopical logistics market scale prediction.Regional logistics demand forecasting for national economy and regional logistics departments set up the scientific development of the future logistics development strategy planning and provide basis for the feasible marketing strategy, and provide reference for macro industrial economic policymaking.In addition, the government can through the regional logistics demand forecasting to assess the overall contributions to the development of logistics industry to the local economy to develop logistics industry development policy, and guide the logistics market resources reasonable use and the optimized configuration.From microscopic perspective, logistics enterprises should be based on the logistics demand forecast reasonable allocation of limited resources, in order to minimize risks and maximize returns.The research status at home and abroadRegional logistics demand forecasting research began in the 1990 s.At the beginning of this decade, Chinese scholars began to study the regional logistics demand forecasting method.After nearly 20 years of development, the regional logistics demand forecasting research have made great progress.Regional logistics demand forecasting method at present, the study found that the main consider regional logistics demand forecasting regression problems.According to the development course and the intelligent level of high and low, can be roughly divided into the following three stages:First stage mainly based on traditional statistical forecast method, it is the turn of the century mainly USES the method of regional logistics demand forecast.The main methods include: input and output model, regression analysis method, the intensity of freight method, elasticity coefficient method, the clustering method, grey theory model, markov chain, space-time many probability model and decision support system, etc.The main characteristic of this kind of method for sequencing and linear data processing, and to construct the model has strong explanatory.With the deepening of the regional logistics demand forecasting research, most of research method is gradually exposed, mainly reflected in the following three aspects: first, the real regional logistics demand data samples were very small and difficult to collect, greatly affecting the effects of prediction methods of validation.Second, in processing high dimensions, containing nonlinear relation, Cheng Fei normal distribution, a time sequence data of regional logistics demand, its effect is not ideal.Third, cannot guarantee that learning and generalization ability, lack of flexibility, and the whole process steps by the regulation, equally to all data on the same treatment regardless of whether the need for these.Behind the two problems have prompted people to consider an artificial intelligence technology in the logistics demand forecasting, in order to improve the performance of the predictive model and improve the prediction accuracy.Artificial intelligence methods used at present scholars mainly is the artificial neural network (ANN) and its modified.The introduction of the artificial intelligence prediction method in traditional forecasting methods into a person's intelligence factors, such as learning and generalization ability of neural network, expert system reasoning rules, etc., so regional logistics demand forecasting technology is a big step forward again., however, can be seen from the previous analysis: first, can not guarantee prediction model is theoretically generalization ability, which makes for forecasting model, after training for the new demand of logistics data sets no stable prediction effect.Second, in the study sample size is limited, the learning process error prone to converge to local minimum point, learning accuracy is difficult to guarantee;Many learning sample variables, and into the "dimension disaster".Third, mainly rely on empirical risk minimization principle, its important faults: use empirical risk instead of expected risk to choose decision function, andwithout strict proof and case well, just a thought reasonable method;Prone to "learn" problem, which is blind pursuit of the inevitable result of the small error and led to the decrease of the generalization ability.Therefore, in order to further improve the precision and stability of the forecasting model, in recent years, researchers began to explore the forecasting method based on structural risk minimization principle, this makes the regional logistics demand forecasting research into the third stage, namely the forecasting method based on statistical learning theory.The method mainly by Support Vector Machine (Support Vector Machine, SVM) as a representative.After the introduction of prediction method based on statistical learning theory, although the researchers did a lot of regional logistics demand, achieved some results, but these methods have shortcomings of one kind or another, or in the regional logistics demand forecasting for solutions to the problems still exist, these weaknesses or problems are mainly concentrated in the following aspects: first,. In the study sample size is limited, the learning process error prone to converge to local minimum point, learning accuracy is difficult to guarantee;Many learning sample variables, and into the "dimension disaster".Second, mainly rely on empirical risk minimization principle., therefore, cannot guarantee prediction model is theoretically generalization ability, which makes for after training, the prediction model of demand for new logistics data sets no stable prediction effect.Third, the concept of regional logistics in our country introduced shortly, many logistics standard not unified, lack of direct statistical data, prediction research institute USES historical data, instead of only USES the related traffic volume, it affects the forecasting methods of empirical research.Especially in China, most of the literature is just a review of methods, almost no use of the actual forecast analysis on regional logistics data, this makes the study of many lack of practical significance.Fourth, the prediction of regional logistics index selection method, researchers and practitioners rely mainly on industry experience, traditional methods, such as principal component analysis, factor analysis, the lack of new way to study.Fifth, these methods, artificial intelligence methods and support vector machine (SVM) method although has certain advantage in prediction accuracy, but cannot effectively explain the model.The future research directionCan be seen through the above analysis, the goal of future research should be in establishing a unified structure framework, based on the regional logistics demand forecasting model using the new regression method to increase the prediction precision of the prediction model and enhancing prediction model can be interpreted.。