CN101419610B - Information processing device, information processing method - Google Patents

Information processing device, information processing method Download PDF

Info

Publication number
CN101419610B
CN101419610B CN 200810170287 CN200810170287A CN101419610B CN 101419610 B CN101419610 B CN 101419610B CN 200810170287 CN200810170287 CN 200810170287 CN 200810170287 A CN200810170287 A CN 200810170287A CN 101419610 B CN101419610 B CN 101419610B
Authority
CN
China
Prior art keywords
feature amount
expression
feature
target
extraction expression
Prior art date
Application number
CN 200810170287
Other languages
Chinese (zh)
Other versions
CN101419610A (en
Inventor
小林由幸
Original Assignee
索尼株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority to JP2007273418A priority Critical patent/JP4392622B2/en
Priority to JP2007-273416 priority
Priority to JP2007273416A priority patent/JP4392621B2/en
Priority to JP2007-273417 priority
Priority to JP2007-273418 priority
Priority to JP2007273417A priority patent/JP4433323B2/en
Application filed by 索尼株式会社 filed Critical 索尼株式会社
Publication of CN101419610A publication Critical patent/CN101419610A/en
Application granted granted Critical
Publication of CN101419610B publication Critical patent/CN101419610B/en

Links

Abstract

An information processing device for generating a target feature amount computational expression for outputting a target feature amount corresponding to input data, comprising: a feature amount extraction expression list generating unit configured to generate and update a feature amount extraction expression list; a feature amount computing unit configured to input actual data supplied as tutor data to each feature amount extraction expression included in the feature amount extraction expression list to compute multiple feature amounts corresponding to the actual data; a target feature amountcomputational expression generating unit configured to employ the multiple feature amounts, and an existing feature amount corresponding to the actual data supplied as tutor data for the same rank togenerate the target feature amount computational expression by machine learning; and an evaluation value computing unit configured to compute the evaluation value of each feature amount extraction expression included in the feature amount extraction expression list.

Description

信息处理设备和信息处理方法 The information processing apparatus and information processing method

[0001] 相关申请的交叉引用 CROSS [0001] REFERENCE TO RELATED APPLICATIONS

[0002] 本发明包含与通过援引将完整内容结合于此的、全部于2007年10月22日向日本专利局提交的日本专利申请JP 2007-273417、日本专利申请JP 2007-273416和日本专利申请JP 2007-27;3418有关的主题内容。 [0002] The present invention comprises the entire contents of which is incorporated herein by reference, filed in the Japanese Patent all Japanese Patent Office on October 22, 2007 Application JP 2007-273417, Japanese Patent Application JP 2007-273416 and Japanese Patent Application JP 2007-27; 3418 related to the subject matter.

[0003] 技术领域 [0003] Technical Field

[0004] 本发明涉及一种信息处理设备、信息处理方法和程序,并且具体地涉及一种实现自动构造特征量计算算法、由此可以计算内容数据如例如音乐数据的特征量的信息处理设备、信息处理方法和程序。 [0004] The present invention relates to an information processing apparatus, information processing method and a program, and particularly relates to an automatic configuration feature quantity calculation algorithm, the contents data can be calculated as the feature amount, for example, music data information processing apparatus, information processing method and a program.

[0005] 背景技术 [0005] BACKGROUND OF THE INVENTION

[0006] 迄今为止已经提出一种利用遗传搜索方法的方法(例如国际公开号W02007/049641)和一种不利用遗传搜索方法的方法(例如美国专利申请公开号US 2004/0181401A1)作为一种用于自动构造特征量计算算法、由此可以输出诸如音乐数据、图像数据等输入数据的特征量(在输入数据是音乐数据的情况下为速度、亮度、生动度)的发明。 [0006] To date a method of using the genetic search method has been proposed (for example, W02007 / 049641 International Publication No.) and a non-search method using genetic methods (such as US Patent Application Publication No. US 2004 / 0181401A1) used as a calculation algorithm for automatic configuration feature amount, whereby the input data can be outputted as the feature quantity of music data, image data and the like (in the case where the input data is music data speed, brightness, liveness) of the present invention.

[0007] 发明内容 [0007] SUMMARY OF THE INVENTION

[0008] 然而,通过相关技术自动构造的特征量计算算法通常包括与人工构造的特征量计算算法相比冗余的算术运算,因而用于获得与输入数据对应的特征量的算术运算所必需的时间在一些情况下延长。 [0008] However, the related art constructed automatically by the feature quantity calculation algorithm generally includes a feature amount calculating artificially constructed redundant arithmetic comparison algorithm, thereby obtaining a feature quantity corresponding to the input data of the arithmetic operation required prolonged in some cases.

[0009] 另外,在无需使用将要开发的特征量计算算法执行计算的情况下即可获得的现有特征量之中,对于构造特征量计算算法而言有效的值得考虑的特征量在特征量计算算法开发者之中已经是公知的,但是目前为止尚未提出用于构造特征量计算算法的方法。 [0009] Further, in the conventional case where the feature quantity of the feature quantity to be developed without the use of the algorithm of the calculation is performed can be obtained, the configuration for the feature quantity calculation algorithm is worth considering effective amount calculating feature quantity of the feature among algorithm developers it is already known, but so far not been proposed a method configured for the feature quantity calculation algorithm.

[0010] 注意下文在本说明书中,无需使用将要自动构造的特征量计算算法执行计算即可获得的现有特征量将称为现有特征量。 [0010] Note that hereinafter in the present specification, calculation is performed to be automatically configured without using the feature quantity calculation algorithm prior to the feature amount obtained will be referred to existing feature amount. 另一方面,将要通过使用特征量计算算法获得的特征量将称为目标特征量, On the other hand, to be calculated by using the feature quantities obtained by the algorithm is called the target feature amount,

[0011] 对于自动构造特征量计算算法、由此通过利用与输入数据对应的现有特征量也可以计算与输入数据对应的目标特征量已经有公认的需求。 [0011] The calculation algorithm for automatic configuration feature amount, the feature amount thus by using the existing input data corresponding to the input data may be calculated with the corresponding target feature amount has been recognized need.

[0012] 作为本发明的一个实施例,根据本发明一个实施例的一种信息处理设备,用于获取输入数据和与输入数据对应的现有特征量作为输入、并且生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该信息处理设备包括:特征量提取表达式列表生成单元,配置成将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于由多个运算符组成的多个特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成包括所述多个特征量提取表达式的特征量提取表达式列表;特征量计算单元,配置成将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量;目标特征量计算表达式生成单元,配置成同等地利用所计算的与实 [0012] As one embodiment of the present invention, an information processing apparatus according to an embodiment of the present invention, the input data and the input data corresponding to the conventional feature amount for obtaining as input and generates an output with the input data target feature amount corresponding to the target feature amount calculation expression, the information processing apparatus comprising: feature amount extraction expression list generating unit, configured to extract a plurality of feature amounts included in the expression list generation feature quantity extraction expression front gene, extraction expression evaluation value by using a plurality of feature amounts based on a plurality of operators composed before the genetic algorithm to update the feature amount extraction expression list generation for generating a plurality of feature amounts of the feature amount extraction expression extraction expression list; feature amount calculation unit configured to input to the feature quantity of each feature amount extraction expression list included in the plurality of feature extraction expression to calculate the actual data corresponding to the actual amount of data supplied from the teaching data ; target feature quantity calculation expression generating unit configured to equally using the calculated real and 际数据对应的多个特征量和与作为教导数据供应的实际数据对应的现有特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式;以及评价值计算单元, 配置成计算在特征量提取表达式列表中包括的每个特征量提取表达式的评价值。 Machine learning actual data corresponding to a plurality of feature values ​​and the actual data as teaching data supplied from the feature quantity corresponding to the conventional through for estimating the actual data supplied from the teaching data corresponding to the target feature amount generating target feature quantity calculation expression; and an evaluation value calculation unit configured to calculate the feature amount extraction expression list of each feature amount extraction expression includes an evaluation value.

[0013] 目标特征量计算表达式生成单元可以同等地有选择地利用所计算的与实际数据对应的多个特征量中的一些特征量和与作为教导数据供应的实际数据对应的多个现有特征量中的一些特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式。 [0013] certain feature quantity calculation expression generating unit may equally be selectively use some of the calculated feature amount and the actual data corresponding to the plurality of feature amounts and the actual data as teaching data corresponding to a plurality of existing supply Some feature amount by the feature amount is to estimate the actual data for the teaching data supplied from the feature quantity corresponding to the target machine learning to generate the target feature amount calculation expression.

[0014] 评价值计算单元可以基于所计算的与实际数据对应的多个特征量的每个目标特征量计算表达式的贡献率来计算特征量提取表达式列表中包括的所述特征量提取表达式的评价值。 [0014] The evaluation value calculation unit may calculate the contribution rate based on the expression of each target feature amount calculated a plurality of feature amounts corresponding to the actual data to calculate the feature amount extraction expression list feature amount extraction expression includes evaluation value formula.

[0015] 根据本发明一个实施例的一种信息处理设备,用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该信息处理设备包括:特征量提取表达式列表生成单元,配置成将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于由多个运算符组成的多个特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成包括所述多个特征量提取表达式的特征量提取表达式列表;特征量计算单元,配置成将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量、并且也测量相应特征量提取表达式的平均计算时间;目标特征量计算表达式生成单元,配置成利用所计算的与实际数据对应的多个特征量以通过用于估计与 [0015] An information processing apparatus according to one embodiment of the present invention, the target feature amount for generating the target feature amount corresponding to the output of the input calculation expression data, the information processing apparatus comprising: feature amount extraction expression list generating unit, configured to extract a plurality of first feature quantity generation feature amount extraction expression list including gene expression, genetic extraction expression evaluation value based on the plurality of feature amounts of a plurality of operators by using the composition before generation algorithm updates the feature amount extraction expression list comprises generating a plurality of feature amount extraction expression list of feature amount extraction expression; feature amount calculation unit configured to input the actual data as teaching data supplied to the feature quantity each feature amount extraction expression list including extraction expression to calculate the actual data corresponding to a plurality of feature amounts, and also measuring the corresponding feature amount extraction expression average computation time; target feature amount calculation expression generating unit configured as calculated using the actual data corresponding to a plurality of feature amounts for estimating by 作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式;以及评价值计算单元,配置成计算在特征量提取表达式列表中包括的每个特征量提取表达式的评价值、并且也基于相应特征量提取表达式的平均计算时间来校正所计算的评价值。 As a practical teaching data corresponding to the target data supplied from the feature amount generating machine learning target feature quantity calculation expressions; and an evaluation value calculation unit configured to calculate each feature amount extraction expression list included in the feature quantity extraction expression evaluation value, and also calculates the time based on the average feature amount extraction expressions corresponding to the corrected evaluation value calculation.

[0016] 目标特征量计算表达式生成单元可以有选择地利用所计算的与实际数据对应的多个特征量中的一些特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式。 [0016] certain feature quantity calculation expression generating unit may be selectively some of the plurality of feature amounts corresponding to the actual data and the feature amount calculated in the through use for estimating the actual data supplied from the teaching data corresponding to the target feature the amount of the target machine learning to generate the feature quantity calculation expression.

[0017] 目标特征量计算表达式生成单元可以基于对应特征量提取表达式的平均计算时间有选择地利用所计算的与实际数据对应的多个特征量中的一些特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式。 [0017] certain feature quantity calculation expression generating unit may correspond to the feature amount extraction expression based on an average feature amount computation time some more feature amount corresponding to the actual data using the calculated selectively in order for estimating by as a practical teaching data corresponding to the target data supplied from the feature amount generating machine learning target feature amount calculation expression.

[0018] 根据本发明一个实施例的一种信息处理设备,用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该信息处理设备包括:特征量提取表达式列表生成单元,配置成将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于由多个运算符组成的多个特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成包括所述多个特征量提取表达式的特征量提取表达式列表;特征量计算单元,配置成将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量;目标特征量计算表达式生成单元,配置成利用所计算的与实际数据对应的多个特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机 [0018] An information processing apparatus according to one embodiment of the present invention, the target feature amount for generating the target feature amount corresponding to the output of the input calculation expression data, the information processing apparatus comprising: feature amount extraction expression list generating unit, configured to extract a plurality of first feature quantity generation feature amount extraction expression list including gene expression, genetic extraction expression evaluation value based on the plurality of feature amounts of a plurality of operators by using the composition before generation algorithm updates the feature amount extraction expression list comprises generating a plurality of feature amount extraction expression list of feature amount extraction expression; feature amount calculation unit configured to input the actual data as teaching data supplied to the feature quantity each feature amount extraction expression list including a plurality of feature extraction expression to calculate the actual data corresponding to an amount; target feature amount calculation expression generating unit, a plurality of feature amounts corresponding to the actual data to be calculated using machine through for estimating the actual data supplied from the teaching data corresponding to the target feature quantity 器学习来生成目标特征量计算表达式;评价值计算单元,配置成计算在特征量提取表达式列表中包括的每个特征量提取表达式的评价值;以及优化单元,配置成优化在最后一代特征量提取表达式列表中包括的多个特征量提取表达式中的每个特征量提取表达式。 Generating a learning target feature quantity calculation expressions; evaluation value calculating unit configured to calculate each feature amount extraction expressions included in the list of the extraction expression evaluation value feature quantity; and an optimizing unit configured to optimize the generation of the final a plurality of feature amount extraction expression list of feature amounts included in each extracted feature amount extraction expression expression.

[0019] 优化单元可以包括:特征量提取表达式优化单元,配置成通过以下方式来优化在最后一代特征量提取表达式列表中包括的每个特征量提取表达式:即根据在最后一代特征量提取表达式列表中包括的相应特征量提取表达式检测表示了预先登记的冗余运算符的组合的优化模式并且删除运算符或者替换为算术负荷较小的运算符。 [0019] The optimization unit may include: a feature amount extraction expression optimizing unit configured to optimize each feature amount extraction expression lists in the final generation of the feature amount extraction expressions included in the following manner: i.e. the amount in the last generation feature corresponding to the feature amount extraction expression list extraction expression comprising detecting a combined optimization model registered in advance and deletes redundant operator is an arithmetic operator or alternatively smaller load operator.

[0020] 优化单元可以包括:特征量提取表达式优化单元,配置成:变形在最后一代特征量提取表达式列表中包括的每个特征量提取表达式以生成多个优化候选表达式;向多个生成的优化候选表达式之中的以下优化候选表达式赋予优良评价,获得的该优化候选表达式的输出具有与作为变形源的特征量提取表达式的输出的高相关度、并且该优化候选表达式的计算时间较短;将多个生成的优化候选表达式作为基因,利用基于优化候选表达式的评价的遗传算法以更新多个生成的优化候选表达式;以及将具有最优良评价的优化候选表达式最终确定为在最后一代特征量提取表达式列表中包括的相应特征量提取表达式的优化结果。 [0020] The optimization unit may include: a feature amount extraction expression optimizing unit is configured to: extract each modified feature amount extraction expression list comprising an expression optimized to generate a plurality of candidate expressions in the last generation of the feature amount; to multiple the following expression gives good candidates for optimization evaluation, optimization of the expression of the candidate having outputs the obtained feature amount extraction expression as a modification of the source output by a high degree of correlation, and the optimization generating a candidate from among the candidates for optimization expression calculating a short time expression; generating a plurality of candidates for optimization of gene expression, genetic optimization algorithm based on evaluating a candidate expression to update the plurality of candidate expressions to generate optimized; and the most excellent evaluation in the optimization the final expression determined candidate corresponding feature amount included in the last generation of the feature amount extraction expression list extraction expression optimization results.

[0021] 优化单元可以包括:重构单元,配置成利用所优化的特征量提取表达式以重构与最后一代特征量提取表达式列表对应生成的目标特征量计算表达式。 [0021] The optimization unit may include: a reconstruction unit, configured using the optimization feature quantity extraction expression to reconstruct the last generation of the feature amount extraction expression list corresponding to the generated object feature amount calculation expression.

[0022] 根据本发明一个实施例的一种信息处理方法,用于获取输入数据和与输入数据对应的现有特征量作为输入、并且生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该信息处理方法包括:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量;同等地利用所计算的与实际数据对应的多个特征量和与作为教导数据供应的实际数据对应的现有特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式;计算在特征量提取表达式列表中包括的每个特征量提取表达式的评价值;以及将在前一代特征量提取表达式列表中包括 [0022] An information processing method according to an embodiment of the present invention, for obtaining the input data and the conventional feature amount corresponding to the input data as input, and generates a target characteristic of the output to the input data corresponding to the target feature quantity calculation expressions, the information processing method comprising: randomly generating a plurality of feature amounts of a plurality of operators composed feature amount extraction expression list extraction expression; the actual data inputted as the teaching data supplied to the feature amount extraction each feature amount extraction expression list including expression to calculate the actual data corresponding to a plurality of feature amounts; equally using the calculated and the actual data corresponding to the plurality of feature amounts and the actual data as teaching data supplied from the corresponding conventional machine learning by the feature quantity to the actual data for estimating the teaching data supplied from the feature quantity corresponding to the target to generate a target feature value calculation expressions; calculated for each feature amount extraction expression list included in the feature quantity extraction expression evaluation value; and the feature amount extraction expression list of the preceding generation comprises 的多个特征量提取表达式作为基因,通过使用基于特征量提取表达式的评价值的遗传算法来更新前一代特征量提取表达式列表来生成下一代特征量提取表达式列表。 A plurality of feature amount extraction for gene expression, genetic algorithms extraction expressions based on the evaluation value feature quantity by using the pre-update generation feature quantity extraction expression list of the next generation feature quantity extraction expression list.

[0023] 根据本发明一个实施例的一种信息处理方法,用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该信息处理方法包括:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量、并且也测量相应特征量提取表达式的平均计算时间;利用所计算的与实际数据对应的多个特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式;计算在特征量提取表达式列表中包括的每个特征量提取表达式的评价值、并且也基于相应特征量提取表达式的平均计算时间来校正所计算的评价值;以及将在前一代特征量提取表达式列表 [0023] An information processing method according to an embodiment of the present invention, the target feature amount for generating the target feature amount corresponding to the output of the input calculation expression data, the information processing method comprising: generating a plurality of random consisting of a plurality of feature amount extraction expression operators feature amount extraction expression list; input to each of the feature amount extraction expression list included in the feature amount extraction expression supplied as actual data to calculate the teaching data and actual data corresponding to the plurality of feature amounts, and also measuring the corresponding feature amount extraction expression average computation time; using the calculated and the actual data corresponding to a plurality of feature amounts by the actual data for estimating the teaching data supplied from the corresponding target feature amount generating machine learning target feature amount calculation expression; calculating for each feature amount extraction expression list including extraction expression evaluation value, and also based on the corresponding feature amount extraction expression average feature amount computation time correcting the calculated evaluation value; and the feature amount extraction expression list of the preceding generation 中包括的多个特征量提取表达式作为基因,通过使用基于特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成下一代特征量提取表达式列表。 The plurality of feature amounts included in the expression for gene extraction, the extraction update algorithm before genetic expression evaluation value generation feature amount extraction expression list of the next generation is generated based on the feature amount extraction expression list by using the feature quantity.

[0024] 根据本发明一个实施例的一种信息处理方法,用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该信息处理方法包括:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量;利用所计算的与实际数据对应的多个特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式;计算在特征量提取表达式列表中包括的每个特征量提取表达式的评价值;将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来 [0024] An information processing method according to an embodiment of the present invention, the target feature amount for generating the target feature amount corresponding to the output of the input calculation expression data, the information processing method comprising: generating a plurality of random consisting of a plurality of feature amount extraction expression operators feature amount extraction expression list; input to each of the feature amount extraction expression list included in the feature amount extraction expression supplied as actual data to calculate the teaching data and actual data corresponding to the plurality of feature amounts; machine learning wherein a plurality of the actual data corresponding to an amount by using the calculated to estimate the actual data for the teaching data supplied from the feature quantity corresponding to the target to generate a target feature amount calculation expression ; calculated for each feature amount extraction expression list are included in the feature amount extraction expression evaluation value; and the plurality of first feature quantity generation extracting feature amount extraction expression list including gene expression, characterized by based amount extraction expression evaluation value of the genetic algorithm generation feature quantity extraction expression list before updating 生成下一代特征量提取表达式列表;以及优化在最后一代特征量提取表达式列表中包括的多个特征量提取表达式中的每个特征量提取表达式。 Generating a feature amount extraction expression list of the next generation; and optimize each feature amount of the plurality of feature amounts included in the last generation of the feature amount extraction expression list extraction expression extraction expressions.

[0025] 根据本发明一个实施例的一种用于控制信息处理设备并且使信息处理设备的计算机执行处理的程序,该信息处理设备用于获取输入数据和与输入数据对应的现有特征量作为输入、并且生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该处理包括以下步骤:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量;同等地利用所计算的与实际数据对应的多个特征量和与作为教导数据供应的实际数据对应的现有特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式;计算在特征量提取表达式列表中包括的每个特征 [0025] According to one embodiment of the present invention is a method for controlling the information processing apparatus and a program causing a computer to execute processing of the information processing apparatus, information processing apparatus for acquiring the input data and the corresponding current feature quantity of input data as input, and generates a target output characteristic corresponding to the input data and the target feature amount calculation expression, the process comprising the steps of: randomly generating a plurality of feature amounts of a plurality of operators composed of extracting feature quantity extraction expression expression list; teaching data as the actual input data supplied to each feature amount extraction expression list included in the feature amount extraction expression to calculate the actual data corresponding to a plurality of feature amounts; equally calculated using the machine learning the actual data corresponding to a plurality of feature values ​​and the actual data as teaching data supplied from the feature quantity corresponding to the conventional through for estimating the actual data supplied from the teaching data corresponding to the target feature amount generating target feature quantity calculation expression; calculating each feature in the feature amount extraction expressions included in the list 量提取表达式的评价值;以及将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成下一代特征量提取表达式列表。 Amount extraction expression evaluation value; and extracting a plurality of feature amount extraction expressions included in the list for gene expression previous generation feature quantity extraction algorithm updates the previous generation genetic characteristics expression evaluation value based on the feature quantity by using amount extraction expression list of the next generation feature quantity extraction expression list.

[0026] 根据本发明一个实施例的一种用于控制信息处理设备并且使信息处理设备的计算机执行处理的程序,该信息处理设备用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该处理包括以下步骤:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量、 并且也测量相应特征量提取表达式的平均计算时间;利用所计算的与实际数据对应的多个特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式;计算在特征量提取表达式列表中包括的每个特征量提取表达式的评价值、并且也基于相应特征量提取表达式的平均 [0026] The program for controlling the information processing apparatus and causing the computer to execute the information processing apparatus according to one embodiment of the present invention, the information processing apparatus for generating output data corresponding to the target input characteristic amount target feature amount calculation expression, the process comprising the steps of: randomly generating a plurality of feature amounts of a plurality of operators composed feature amount extraction expression list extraction expression; the actual data inputted as the teaching data supplied to the feature quantity each feature amount extraction expression list including extraction expression to calculate the actual data corresponding to a plurality of feature amounts, and also measuring the respective time average calculating feature amount extraction expression; using the calculated and the actual data corresponding to the plurality feature amount by estimating the actual data for the teaching data supplied from the feature quantity corresponding to the target machine learning to generate the target feature amount calculation expression; calculating for each feature amount extraction expression list included in the feature amount extraction expression type evaluation value, and also based on the average feature amount extraction expressions corresponding 算时间来校正所计算的评价值; 以及将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成下一代特征量提取表达式列表。 And before the extracted plurality of feature amounts included in the expression list generation feature quantity extraction expression for gene first, extracting genetic expression evaluation value update algorithm based on the feature quantity by the use; calculation time correcting the calculated evaluation value generation feature quantity extraction expression list to generate the next generation of feature amount extraction expression list.

[0027] 根据本发明一个实施例的一种用于控制信息处理设备并且使信息处理设备的计算机执行处理的程序,该信息处理设备用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,该处理包括以下步骤:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的每个特征量提取表达式以计算与实际数据对应的多个特征量; 利用所计算的与实际数据对应的多个特征量以通过用于估计与作为教导数据供应的实际数据对应的目标特征量的机器学习来生成目标特征量计算表达式;计算在特征量提取表达式列表中包括的每个特征量提取表达式的评价值;将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于特 [0027] The program for controlling the information processing apparatus and causing the computer to execute the information processing apparatus according to one embodiment of the present invention, the information processing apparatus for generating output data corresponding to the target input characteristic amount target feature amount calculation expression, the process comprising the steps of: randomly generating a plurality of feature amounts of a plurality of operators composed feature amount extraction expression list extraction expression; the actual data inputted as the teaching data supplied to the feature quantity each feature amount extraction expression list including extraction expression to calculate the actual data corresponding to a plurality of feature amounts; using the calculated and the actual data corresponding to a plurality of feature amounts for estimating by supplying the teaching data, as target feature amount corresponding to the actual data to generate a target machine learning feature quantity calculation expression; calculating for each feature amount extraction expressions included in the list of the extraction expression evaluation value feature quantity; the first feature amount extraction expression generation the list includes a plurality of feature amount extraction expression gene, by using Laid-based 量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成下一代特征量提取表达式列表;以及优化在最后一代特征量提取表达式列表中包括的多个特征量提取表达式中的每个特征量提取表达式。 Amount extraction expression evaluation value generation before the genetic algorithm to update the feature amount extraction expression list of the next generation feature quantity extraction expression list; and optimizing a plurality of feature amounts included in the last generation of the feature amount extraction expression list is extracted Expression wherein each feature amount extraction expression.

[0028] 利用本发明的一个实施例,生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表,将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的相应特征量提取表达式以计算与实际数据对应的多个特征量。 [0028] With the present invention one embodiment includes a plurality of feature amounts generated by a plurality of operators composed feature amount extraction expression extraction expression list, the input to the feature amount extraction expression as the actual data supplied from the teaching data the corresponding feature quantity extraction formula included in the list of expressions to calculate the actual data corresponding to the plurality of feature amounts. 另外,通过机器学习来生成目标特征量计算表达式,其中同等地利用所计算的与实际数据对应的多个特征量和与作为教导数据供应的实际数据对应的现有特征量以估计与作为教导数据供应的实际数据对应的目标特征量。 Further, the target feature is generated by a machine learning amount calculating expressions, which are equally calculated using the actual data corresponding to a plurality of feature values ​​and the actual data as teaching data supplied from the feature amount corresponding to the prior teachings as to estimate target feature amount data corresponding to actual supply. 另外,计算在特征量提取表达式列表中包括的相应特征量提取表达式的评价值,并且将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成下一代特征量提取表达式列表。 Additionally, computing respective feature amounts extracted in the feature amount extraction expression list includes expression evaluation value, and extracts a plurality of feature amount extraction expressions included in the list for gene expression previous generation feature quantity, based by before the feature amount extraction expression evaluation value of genetic algorithm update feature amount extraction expression list generation to the next generation feature quantity extraction expression list.

[0029] 利用本发明的一个实施例,随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表,将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的相应特征量提取表达式以计算与实际数据对应的多个特征量,并且也测量相应特征量提取表达式的平均计算时间。 [0029] With the present invention, one embodiment comprises a plurality of randomly generated by a plurality of feature amount extraction expression operators composed of a feature amount extraction expression lists, as the actual input data supplied to a teaching data extracting feature amount expression corresponding to the feature amount extraction expression list comprises to calculate the actual data corresponding to a plurality of feature amounts, and also measuring the respective time average calculating feature amount extraction expressions. 另外,通过机器学习来生成目标特征量计算表达式,其中利用所计算的与实际数据对应的多个特征量以估计与作为教导数据供应的实际数据对应的目标特征量,计算在特征量提取表达式列表中包括的相应特征量提取表达式的评价值,并且也基于相应特征量提取表达式的平均计算时间来校正所计算的评价值。 Further, generated by a machine learning target feature amount calculation expression, wherein using the calculated and the actual data corresponding to the plurality of feature amounts as to estimate the actual data supplied from the teaching data corresponding to the target feature amount, the feature amount calculated in the extracted expression the corresponding feature quantity extraction formula list includes expression evaluation value, and also calculates the time based on the average feature amount extraction expressions corresponding to the corrected evaluation value calculation. 另外,将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成下一代特征量提取表达式列表。 Further, the extracted feature quantity generation of a plurality of first feature amount extraction expression list including gene expression, genetic extraction expression evaluation value update algorithm of the previous generation feature quantity based on the feature amount extraction expression list is generated by using next-generation feature quantity extraction expression list.

[0030] 利用本发明的一个实施例,随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表,将作为教导数据供应的实际数据输入到在特征量提取表达式列表中包括的相应特征量提取表达式以计算与实际数据对应的多个特征量。 [0030] With the present invention, one embodiment comprises a plurality of randomly generated by a plurality of feature amount extraction expression operators composed of a feature amount extraction expression lists, as the actual input data supplied to a teaching data extracting feature amount expression corresponding to the feature amount extraction expression list includes calculating a plurality of feature amounts corresponding to the actual data. 另外,通过机器学习来生成目标特征量计算表达式,其中利用所计算的与实际数据对应的多个特征量以估计与作为教导数据供应的实际数据对应的目标特征量,并且计算在特征量提取表达式列表中包括的相应特征量提取表达式的评价值。 Further, generated by a machine learning target feature amount calculation expression, wherein using the calculated and the actual data corresponding to the plurality of feature amounts as to estimate the actual data supplied from the teaching data corresponding to the target feature amount, and calculates a feature amount extraction expression list comprising a respective feature amount extraction expression evaluation value. 随后,将在前一代特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于特征量提取表达式的评价值的遗传算法更新前一代特征量提取表达式列表来生成下一代特征量提取表达式列表。 Subsequently, the extracted feature quantity generation of a plurality of first feature amount extraction expression list including gene expression, based on the feature quantity extraction algorithm before updating genetic expression evaluation value by using the feature amount extraction expression list generation to generate next-generation feature quantity extraction expression list. 另外,优化在最后一代特征量提取表达式列表中包括的多个特征量提取表达式。 Further, optimization of the plurality of extracted feature amount extraction expressions included in the list in the final expression generation feature quantity.

[0031] 根据本发明的一个实施例,可以自动构造特征量计算算法,由此甚至可以利用输入数据的现有特征量来计算与输入数据对应的目标特征量。 [0031] According to one embodiment of the present invention, the feature quantity calculation algorithm configured automatically, whereby even existing feature amount of input data corresponding to the input data to calculate a target feature amount.

[0032] 根据本发明的一个实施例,可以自动构造特征量计算算法,由此可以通过对计算时间施加限制来来计算与输入数据对应的目标特征量。 [0032] In accordance with one embodiment of the present invention may be configured to automatically feature quantity calculation algorithm, thereby to calculate by calculating the time limit is applied to the input data corresponding to the target feature amount.

[0033] 根据本发明的一个实施例,可以自动构造特征量计算算法,由此可以无冗余地计算与输入数据对应的目标特征量。 [0033] According to one embodiment of the present invention, the feature quantity calculation algorithm configured automatically, whereby non-redundant input data corresponding to the calculated target characteristic quantity. 附图说明 BRIEF DESCRIPTION

[0034] 图1是用于描述通过已经应用了本发明一个实施例的目标特征量计算表达式构造系统而生成的特征量计算表达式的图; [0034] FIG. 1 is a calculation expression described by the present invention is applied has a feature amount embodiment target feature amount calculation expression system configured to generate FIG embodiment;

[0035] 图2是图示了教导数据的数据结构的图; [0035] FIG. 2 is a diagram illustrating a data structure of teaching data;

[0036] 图3是示出已经应用了本发明一个实施例的目标特征量计算表达式构造系统的配置例子的框图; [0036] FIG. 3 is a block diagram illustrating the target has feature quantity calculation configuration example of embodiment of the expression of a configuration of a system embodiment of the present invention;

[0037] 图4是图示了特征量提取表达式的例子的图; [0037] FIG. 4 is a diagram illustrating an example of the feature amount extraction expressions FIG;

[0038] 图5是用于描述特征量提取表达式的结构的图; [0038] FIG. 5 is a view showing the structure described feature amount extraction expressions;

[0039] 图6是图示了特征量提取表达式列表的例子的图; [0039] FIG. 6 is a diagram illustrating the feature amount extraction expression list of the example of FIG;

[0040] 图7是用于描述遗传算法的图; [0040] FIG. 7 is a diagram for describing a genetic algorithm;

[0041] 图8是用于描述已经应用了本发明一个实施例的目标特征量计算表达式构造系统的操作的流程图; [0041] FIG. 8 is for describing the present invention has been applied a flowchart of the target feature amount calculation operation expressions embodiment of a system configuration of embodiment;

[0042] 图9是用于具体描述图8中所示步骤S4的流程图; [0042] FIG. 9 is a flowchart specifically described steps shown in FIG. 8 S4 for;

[0043] 图10是图示了选择表组的例子的图; [0043] FIG. 10 is a diagram illustrating an example of a group selection table of FIG;

[0044] 图11是用于具体描述图8中所示步骤SlO的流程图; [0044] FIG. 11 is a flowchart shown in FIG. 8 SlO specifically described for;

[0045] 图12是用于具体描述图11中所示步骤S42的流程图;以及 [0045] FIG. 12 is a flowchart for specifically describing the steps shown in FIG. 11 S42; and

[0046] 图13是图示了计算机的配置例子的框图。 [0046] FIG. 13 is a block diagram illustrating a configuration example of a computer.

具体实施方式 Detailed ways

[0047] 下文将参照附图关于已经应用本发明的具体实施例具体地进行描述。 [0047] Hereinafter will be described with reference to the drawings in particular specific embodiments of the present invention has been applied on.

[0048] 已经应用本发明一个实施例的目标特征量计算表达式构造系统10(图3)利用将要供应的多个教导数据通过机器学习来生成目标特征量计算表达式1,其中如图1中所示, 获取输入数据C和与之对应的多个现有特征量Fl。 Target feature amount plurality of teaching data [0048] The present invention has been applied to one embodiment of the calculation expression construction system 10 (FIG. 3) to be supplied using the target feature value is generated by a machine learning calculation expression 1, wherein 1 in FIG. shown, input data is captured and C corresponding to the plurality of feature amounts prior Fl. 、F2。 , F2. 直至而。 And until. 作为输入,而输出与相关输入数据对应的多个特征量I中的每个特征量。 As an input, and each feature amount corresponding to the input data associated with the plurality of feature amounts I.

[0049] 图2示出了教导数据的数据结构。 [0049] FIG. 2 shows a data structure of teaching data. 也就是说,教导数据TiG = 1、2直至L)由作为与输入数据C同一种数据的实际数据队、与实际数据Di对应的多个现有特征量Fli至!^、 以及与实际数据Di对应的多个目标特征量Ili至Iki组成。 That is, until the teaching data TiG = 1,2 L) as the actual data and the input data C of the team of the same data, a plurality of feature amounts and the actual current data Di corresponding to Fli! ^, And the actual data Di a plurality of feature amounts corresponding to the target composition Ili to Iki.

[0050] 现有特征量? [0050] existing feature amount? 、至!^是以下值,这些值表示了使用现有方法将要从实际数据Di中检测的实际数据Di的特征。 To! ^ The following values, which represent the characteristics of the conventional method using the real data Di from the actual detected data Di. 目标特征量Ili至Iki是以下值,这些值表示了使用现有方法无法从实际数据Di中检测的实际数据Di的特征,例如通过对使多人监视实际数据Di而获得的印象进行数字化所确定的值。 Target feature amount Ili to Iki following values, which illustrates the use of the conventional methods can not detect the actual data from the actual data Di Di features, such as determined by the digitized impression that the people monitoring the actual data Di obtained value.

[0051 ] 如图2中所示,在有k类目标特征量的情况下,目标特征量计算表达式构造系统10 生成k个目标特征量计算表达式。 As shown in Figure [0051] 2, there are k classes in the target feature amount, the feature amount calculating target expression system 10 configured to generate a target k-th feature quantity calculation expression.

[0052] 注意只要输入数据C是多维数据,那么输入数据C的类型就是任意的。 [0052] Note that as long as the input data C are multidimensional data, the type of input data C is arbitrary. 例如,具有时间维度和声道维度的音乐数据、具有维度X、维度Y和像素维度的图像数据、通过向图像数据添加时间维度而获得的移动图像数据等可以用作输入数据C。 For example, music data having a time dimension and a dimension of the channel, having dimensions X, Y dimensions of the image data and the pixel dimensions, moving image data by adding the time dimension to the image data obtained may be used as input data C.

[0053] 注意在以下描述中将描述利用音乐数据作为输入数据C的例子。 [0053] Note that as described using the example of the input data is music data C in the following description. 与音乐数据对应的多个现有特征量的例子包括节奏、速度和节奏波动。 Examples of the music data corresponding to the plurality of feature amounts include conventional rhythm, tempo, and rhythm fluctuation. 与音乐数据对应的目标特征量的例子也包括音乐数据的亮度和速度以及乐器的多样性。 And the music data corresponding to the target feature amounts example also includes the diversity of the music data and speed, and the brightness of the instrument.

[0054] 图3图示了已经应用本发明一个实施例的目标特征量计算表达式构造系统10的配置例子。 [0054] FIG. 3 illustrates an application of the present invention has a configuration example of the target feature quantity calculation expression construction system embodiment 10 of FIG. 目标特征量计算表达式构造系统10由以下单元配置而成:特征量提取表达式列表生成单元11,用于生成和更新由多个特征量提取表达式组成的特征量提取表达式列表; 特征量计算单元12,用于将教导数据Ti的实际数据Di代入所生成的相应特征量提取表达式以计算特征量;目标特征量计算表达式生成单元13,用于通过机器学习来生成目标特征量计算表达式、由此可以根据由特征量计算单元12计算的与教导数据Ti对应的特征量以及根据教导数据Ti的现有特征量Fli至!^i来估计教导数据Ti的目标特征量Ili至Ui,以及还用于计算每个特征量提取表达式的评价值;以及优化单元15,用于优化最终已经更新的最后一代特征量提取表达式列表和目标特征量计算表达式。 Target feature amount calculation expression system 10 is configured by the configuration unit together: feature amount extraction expression list generating unit 11 for generating and updating the feature amount extracted by the expression consisting of a plurality of feature amount extraction expression list; feature amount corresponding feature quantity calculation unit 12, the teaching data for the actual data Di Ti substitutes the generated feature amount extraction expression to calculate; target feature amount calculation expression generating unit 13 for calculating a target feature value generating machine learning expression, thereby Ti corresponding to the feature amount calculated by the calculation unit 12 and the teaching data and the feature amount data in accordance with the teachings of the prior feature quantity Ti to Fli! ^ i to estimate the target characteristic quantity of Ti to the teaching data Ili Ui , and also for calculating a feature amount extraction expression for each evaluation value; and an optimizing unit 15 configured to optimize the final last generation has been updated feature amount extraction expression list of the target and the feature amount calculation expression.

[0055] 特征量提取表达式列表生成单元11随机生成组成第一代特征量提取表达式列表的多个特征量提取表达式并且将这些表达式输出到特征量计算单元12。 [0055] The feature amount extraction expression list generating unit 11 composed of a plurality of randomly generated features a first generation feature quantity extraction expression amount extraction expression list and outputs these expressions to the feature amount calculation unit 12.

[0056] 现在将参照图4关于特征量提取表达式列表生成单元11将要生成的特征量提取表达式进行描述。 [0056] FIG. 4 will now feature amount of the feature amount extraction expression list generating unit 11 to be generated be described with reference extraction expression. 图4A至4D分别图示了特征量提取表达式例子。 4A to 4D illustrate an example of the feature amount extraction expressions.

[0057] 利用特征量提取表达式,在左边描述输入数据的类型,而根据将要计算的顺序在输入数据的类型的右边描述一类或者多类运算符。 [0057] using the feature amount extraction expression describing the type of input data on the left, and the order to be calculated according to the description of a class or multi-class operator input data to the right type. 每个运算符适当地包括将要处理的轴并包括参数。 Each operator suitably includes a shaft to be treated and comprising a parameter.

[0058] 运算符的例子包括平均数(Mean)、快速傅立叶变换(FFT)、标准方差(MDev)、 出现率(Ratio)、低通滤波器(LPF)、高通滤波器(HPF)、绝对值(ABS)、平方(Sqr)、平方根(Sqrt)、正规化(Normalize)、微分(Differential)、积分(Integrate)、最大值(MaxIndex)、全体离差(UVariance)和下采样(DownSampling)。 Examples [0058] The operators include Mean (Mean), a fast Fourier transform (FFT), standard deviation (MDev), appearance ratio (Ratio), a low pass filter (LPF), a high pass filter (HPF), an absolute value (ABS), square (Sqr), square root (Sqrt), normalized (Normalize), differential (differential), integral (Integrate), maximum value (MaxIndex), all the deviation (UVariance) and downsampling (downSampling). 注意在一些情况下根据确定的运算符来固定将要处理的轴,因此在这一情况下利用与参数固定的将要处理的轴。 Note that the operator determines in accordance with shaft fixed to be processed, so the use of the fixed shaft and the parameters to be processed in this case in some cases. 另外,在已经确定伴随有参数的运算符情况下,该参数也被确定为已经随机或者预先设置的值。 Furthermore, it has been determined in the accompanying parameter operator, this parameter is also determined to have a value set in advance or random.

[0059] 例如在图4A中所示特征量提取表达式的情况下,12ToneSM是输入数据,而32#Differential,32#MaxIndexU6#LPF_l ;0. 861 和16#UVariance 各自为运算符。 The case where [0059] for example, the feature amount extraction expression is shown in FIG. 4A, 12ToneSM input data, and 32 # Differential, 32 # MaxIndexU6 # LPF_l;. 0 861 16 # UVariance and each operator. 另外, 在相应运算符内的3姊、16#等表示将要处理的轴。 In addition, 3 percent, # 16 in the other axis representing the respective operator to be processed.

[0060] 现在,12ToneSM表示沿着时间轴对单声道PCM(脉码调制声源)波形数据进行音乐间隔分析,48#表示声道轴,3¾表示频率轴和音乐间隔轴,而16#表示时间轴。 [0060] Now, 12TonesM musical interval analyzing representation mono PCM (pulse code modulation sound source) waveform data along the time axis, # indicates the channel axis 48, 3¾ axis represents frequency and the musical interval axis and represented by # 16 timeline. 运算符的0. 861是低通滤波器处理的参数并且例如表示所发送的频率的阈值。 0.861 operator low-pass filter processing parameters and threshold frequency e.g. represented transmitted.

[0061] 注意组成第一代特征量提取表达式列表的相应特征量提取表达式的输入数据类型与输入数据C的类型相同,随机确定运算符的数目和运算符的类型,但是如图5中所示, 进行在生成每个特征量提取表达式的时间上的限制,其中在依次执行与多个运算符对应的算术运算符时,算术运算的拥有维度的数目依次减少,而每个特征量提取表达式的最终算术运算结果变成标量倍数,或者其维度数目变成预定的较小值(例如1、2等)。 [0061] Note that the composition of the respective first feature amount extraction expression list generation feature quantity extraction expression of the same type of input data type of the input data C, and the random number to determine the type of operators operator, but as shown in FIG. 5 shown, each performed when generating the feature amount extraction expressions time limit, wherein a plurality of sequentially performing arithmetic operators corresponding to operator, has a number of dimensions of the arithmetic operation is sequentially reduced, and each feature quantity the final extraction expression scalar arithmetic operation result becomes multiple, or a smaller number of dimensions becomes a predetermined value (e.g. 1 and 2).

[0062] 从图4A至4D中所示例子中可以理解,利用特征量提取表达式计算的特征量没有变成按照现有概念确定为有意义的值,比如关于音乐数据的节奏、关于图像数据的像素直方图等。 [0062] It will be appreciated from the example illustrated in FIGS. 4A to 4D, using the feature amount extraction expression evaluation feature amount does not become as determined in accordance with the conventional concept of a significant value, such as on the rhythm of the music data, image data on the pixel histogram. 也就是说,在简单地将输入数据代入特征量提取表达式时,利用特征量提取表达式计算的特征量可能是算术运算结果。 That is, when simply the input data into feature amount extraction expression, using a feature amount extraction expression may be calculated feature quantity arithmetic operation result.

[0063] 现在如图6中所示,假设特征量提取表达式列表生成单元11所生成的特征量提取表达式列表由m个特征量提取表达式Π至fm组成。 [0063] Now in FIG 6 is assumed characteristic amount extraction expression list generating unit 11 generates the feature amount extraction expression list composed of m feature amount extraction expression Π to fm composition. 作为特征量提取表达式Π至fm的输入数据的WavM是单声道PCM波形数据,而拥有维度是时间轴和声道轴。 As the feature amount extraction expression to the input data Π WavM fm is the mono PCM waveform data, and have the dimensions and the channel axis is a time axis.

[0064] 现在将回到对图3的描述。 [0064] The description will now return to FIG. 3. 特征量提取表达式列表生成单元11通过根据遗传算法(GA)更新前一代特征量提取表达式列表来生成第二代及其之后的特征量提取表达式列表。 Feature amount extraction expression list generation unit 11 generates the feature amount extraction expression list prior to update according to the genetic algorithm (GA) is generated after the second generation feature quantity extraction expression list and.

[0065] 现在“遗传算法”意味着一种用于使用选择处理、相交处理、变异处理和随机生成处理从当前一代基因来生成下一代基因的算法。 [0065] Now "Genetic Algorithm" means a method for using selection processing, intersection processing, handling and random mutation generation processing algorithm to generate the next generation from the current genes gene. 具体而言,将组成特征量提取表达式列表的多个相应特征量提取表达式作为基因,根据组成当前一代特征量提取表达式列表的多个特征量提取表达式的评价值来执行选择处理、相交处理、变异处理和随机生成处理以生成下一代特征量提取表达式列表。 Specifically, the composition of a plurality of respective feature amounts extracted feature amount extraction expression lists for gene expression, extraction expression evaluation value feature quantity depending on the composition of the current generation of the plurality of feature amount extraction expression list selection process is performed, intersection processing, process variation and random generation process to generate the next feature amount extraction expression list.

[0066] 也就是说,例如如图7中所示,利用对组成当前一代特征量提取表达式列表的多个相应特征量提取表达式的选择处理来选择具有高评价值的特征量提取表达式f2。 [0066] That is, as shown in FIG. 7, by using a plurality of respective feature amounts of the composition of the current generation of feature amount extraction expression list extraction expression selection processing to select the feature amount having a high evaluation value extraction expression f2. 利用对组成当前一代特征量提取表达式列表的多个相应特征量提取表达式的相交处理来使具有高评价值的多个特征量提取表达式f2和f5相交(组合)以生成特征量提取表达式,并且在下一代特征量提取表达式列表中包括这一特征量提取表达式。 Use of a plurality of feature amounts corresponding to the composition of the current generation of feature amount extraction expression extraction expression list intersects a plurality of processing having a high evaluation value of the feature amount extraction expression f2 and f5 intersection (composition) to generate a feature amount extraction expression type, and include the extraction expression list feature amount extraction expression in the next generation feature quantity.

[0067] 利用对组成当前一代特征量提取表达式列表的多个相应特征量提取表达式的变异处理来部分地变异(改变)具有高评价值的特征量提取表达式f2以生成特征量提取表达式,并且在下一代特征量提取表达式列表中包括这一特征量提取表达式。 [0067] Variation of expression using the extracted feature amounts corresponding to the plurality of current generation feature quantity consisting extraction expression list processing part variation (changing) having a high evaluation value feature amount extraction expression f2 to generate a feature amount extraction expression type, and include the extraction expression list feature amount extraction expression in the next generation feature quantity. 利用随机操作来随机生成新的特征量提取表达式,并且在下一代特征量提取表达式列表中包括这一新的特征量提取表达式。 Random operation generate a new random feature amount extraction expression, and extraction expression list to include this new feature amount extraction expression in the next generation feature quantity.

[0068] 现在将回到对图3的描述。 [0068] The description will now return to FIG. 3. 特征量计算单元12将供应的教导数据Ti的实际数据Di代入组成从特征量提取表达式列表生成单元11供应的特征量提取表达式列表的相应特征量提取表达式Π至fm以计算关于教导数据Ti的特征量、也测量相应特征量提取表达式f 1至fm的计算所必需的计算时间,并且在通过将不同的L条实际数据Di代入每个特征量提取表达式来执行计算时计算平均计算时间。 Feature amount calculation unit 12 the teaching data supplied from the actual data Di Ti composition is substituted extraction expression list generating unit 11 supplies the feature quantity from the feature amount extraction expression lists of the respective feature amount extraction expressions to compute Π to fm on the teaching data when calculating the average feature amount of Ti, was also measured corresponding feature amount extraction expression f calculates a calculation time necessary to fm, and in the feature quantity by each of the L different data Di substituting actual extraction expression to perform the calculations calculating time. 计算的特征量和计算的平均计算时间供应到目标特征量计算表达式生成单元13。 Feature amount calculation and average computation time is supplied to the calculated target characteristic quantity calculation expression generating unit 13.

[0069] 如上所述,教导数据Ti的数目是L,而组成特征量提取表达式列表的特征量提取表达式的数目是m,因而在特征量计算单元12计算(LXm)个特征量。 [0069] As described above, the number of Ti teaching data is L, and the composition of the feature amount extraction expression lists of the number of feature amount extraction expression is m, and thus the feature amount calculating unit 12 calculates (LXM) feature amounts. 在下文中,通过将教导数据TiG = 1、2直至L)的实际数据Di代入特征量提取表达式fj(j = 1、2直至m)而计算的特征量将称为fj[Tj。 Hereinafter, by the teaching data until TiG = 1,2 L) is substituted for the actual data Di feature amount extraction expression fj (j = 1,2 until m) calculated feature quantity will be referred to fj [Tj.

[0070] 每当从特征量计算单元12供应与当前一代特征量提取表达式列表对应的(LXm) 个特征量fj [Ti]时,目标特征量计算表达式生成单元13利用作为特征量计算单元12的计算结果的(LXm)个特征量fj [TJ、在教导数据Ti中包括的(LXn)个现有特征量Fli至! When [0070] 12 each time calculated from the feature amount extraction expression list of the supply unit corresponds to the current generation feature quantity (LXM) feature amounts fj [Ti], the target characteristic quantity calculation expression generating unit 13 calculates a feature quantity by using means (LXM) feature amount calculation result 12 fj [TJ, (LXn) a conventional feature amount included in the teaching data to Ti in Fli! ^、 在教导数据Ti中包括L个目标特征量Ili通过机器学习(利用特征选择的线性鉴别或者递归)来生成在以下表达式(1)中示出的目标特征量计算表达式,该表达式例如通过在与输入数据C对应的现有特征量F1。 ^, Including L Ili target feature amount generating the target feature quantity (1) shown in the following expression by machine learning (using linear discriminant features selected or a recursive) in the teaching data calculation expression Ti, the expression example, an existing feature amount F1 and the input data C corresponding through. 至而。 And to. 与特征量Π [c]至fm[C]之间的线性耦合来生成目标特征量II。 And feature quantity Π [c] a linear coupling between the [C] fm to the target feature amount generating II. . [0071]目标特征量 Ilc = bo+b! -Flc+b2 -F2C+. · · +bn -Fnc+bn+1 -fl [C]+bn+2 -f2[C]+. · · +bn+mf m[C], . . (1) [0071] certain feature amount Ilc = bo + b! -Flc + b2 -F2C +. · · + Bn -Fnc + bn + 1 -fl [C] + bn + 2 -f2 [C] +. · · + Bn + mf m [C],.. (1)

[0072] 注意在表达式[1]中,h是分节,而VID2直至bn+m是线性耦合系数。 [0072] Note that in the expression [1], h is the section, and VID2 until bn + m is a linear coupling coefficient. 另外,对于在目标特征量计算表达式生成单元13实际生成的目标特征量计算表达式,没有利用所有的现有特征量? Further, the target for the feature amount calculation expression generating unit 13 generates an actual target feature amount calculation expression, without utilizing all existing feature amount? 1。 1. 至而。 And to. 和特征量打[(:]至fm[C]而是有选择地利用这些特征量。在这一情况下,与未利用的现有特征量?1。至而。和特征量fl[C]至fm[C]对应的线性耦合系数设置为零。 And beat feature quantity [(:] to fm [C], but selectively use these feature amounts in this case, the conventional feature amount of unutilized 1 and to the feature quantity fl [C].?.. to [C] corresponds to the linear coupling coefficient is set to zero fm.

[0073] 类似地,也生成目标特征量计算表达式,由此可以分别通过在与输入数据C对应的现有特征量F1。 [0073] Similarly, also generates a target feature amount calculation expression, whereby the feature quantity by the prior input data C corresponding to F1, respectively. 至而。 And to. 与特征量Π [C]至fm[C]之间的线性耦合来生成与输入数据对应的目标特征量12。 And feature quantity Π [C] to a linear coupling between the fm [C] corresponds to the input generates feature amount data of the target 12. 至让。 To make. .

[0074] 因而,在目标特征量计算表达式生成单元13生成k个目标特征量计算表达式。 [0074] Accordingly, the target feature amount calculation expression generating unit 13 generates a target k-th feature quantity calculation expression.

[0075] 随后,在生成的目标特征量计算表达式已经到达所需精确度的情况下,或者在预定指令已经由用户给出的情况下,这时的特征量提取表达式列表作为最后一代特征量提取表达式列表与目标特征量计算表达式一起供应到优化单元15。 [0075] Subsequently, the feature amount calculated in the target generated expression case has reached the required accuracy, or in a case where a predetermined instruction has been given by the user, when the feature amount extraction expression list of the final generation characteristics as amount extraction expression list of the target feature quantity calculation expression supplied to the optimization unit 15 together.

[0076] 另外,目标特征量计算表达式生成单元13利用内置评价值计算单元14来计算组成当前一代特征量提取表达式列表的相应特征量提取表达式的评价值。 [0076] Further, the target characteristic quantity calculation expression generating unit 13 uses the built-in evaluation value calculating means 14 calculates a feature amount corresponding to the composition of the current generation of feature amount extraction expression list extraction expression evaluation value. 也就是说,评价值计算单元14计算k个目标特征量计算表达式各自的每个特征量提取表达式的贡献率,并且将通过合计所计算的k个贡献率而获得的总贡献率确定为组成当前一代特征量提取表达式列表的相应特征量提取表达式的评价值。 That is, the evaluation value calculating unit 14 calculates the k target feature quantity calculation expression for each respective feature amount extraction expressions contribution rate, and the sum obtained by the calculated total contribution of the k contribution rate determined feature amount corresponding to the composition of the current generation of feature amount extraction expression list of the extracted evaluation value of the expression.

[0077] 现在将参照以下表达式(¾对贡献率计算方法进行描述。注意表达式(¾用Xp X2直至xn+m替换现有特征量F1。至而。与特征量Π [C]至fm[C]。 [0077] Now the (¾ calculation method will be described with reference to the contribution of the following expression. Note that expression (Xp X2 until ¾ replaced with xn + m feature amounts prior Fl. To be. Characteristic amount Π [C] to fm [C].

[0078]目标特征量 Ilc = Vb1 · X^b2 · X2+. · · +bn+m · Xn+m. · · (2) [0078] certain feature amount Ilc = Vb1 · X ^ b2 · X2 +. · · + Bn + m · Xn + m. · · (2)

[0079] 利用以下表达式(¾来计算Xm(M = 1、2直至n+m)对表达式(¾计算目标特征量Ilc的贡献率(Xm)。 [0079] is calculated by the following expression (¾ Xm (M = 1,2 until n + m) is calculated contribution ratio of the target characteristic quantity Ilc (Xm of) the expression (¾.

[0080] (Xm) = bM/StDev(Xm) XStDev(11) XCorrel (XM, II). . . (3) [0080] (Xm) = bM / StDev (Xm) XStDev (11) XCorrel (XM, II)... (3)

[0081] 这里,StDev(Xsi)代表已经用于机器学习的LXm(S)的标准方差。 [0081] Here, StDev (Xsi) have been used on behalf LXm (S) of the standard deviation of machine learning.

[0082] StDev(Il)代表已经用于机器学习的教导数据Ti中包括的L个目标特征量Ili的标准方差。 [0082] StDev (Il) has been used on behalf of Ti teaching machine learning data included in the target feature amount L Ili standard deviation.

[0083] Correl (XM, ID代表已经用于机器学习的LXm(S)与教导数据Ti中包括的L个目标特征量Ili之间的皮尔森相关系数。 Pearson correlation coefficient between LXm (S) [0083] Correl (XM, ID Representative has a machine learning data with the teachings included in the L Ti target feature amount Ili.

[0084] 注意如以下表达式(4)中所示,通过将LXm(S)与L个目标特征量Ili之间的协方差除以LXm(S)的标准方差与L个目标特征量Ili的标准方差之间的乘积来计算皮尔森相关系数Correl (XM,II)。 [0084] Note that as shown in the following expression (4) by the covariance between LXm (S) and the L divided by the target feature amount Ili LXm (S) and the standard deviation of the L of the target feature amount Ili the product of the standard deviation between the calculated Pearson correlation coefficient Correl (XM, II).

[0085] Correl (XM,II) = (XM和IIi之间的协方差)/ (XM的标准方差X IIi的标准方差)··· (4) [0085] Correl (XM, II) = (covariance between XM and IIi) / (XM standard deviation standard deviation of X IIi) (4)

[0086] 注意评价值计算单元14可以基于皮尔森相关系数来确定组成当前一代特征量提取表达式列表的相应特征量提取表达式fl至fn的评价值,而不是如上所述基于作为目标特征量计算表达式的相应特征量提取表达式Π至fm的输出值的特征量Π []至fm[]的贡献率来确定评价值。 [0086] Note that the evaluation value calculating means 14 based on the Pearson correlation coefficient to determine the composition of the evaluation value corresponding to this feature amount extraction expression list generation feature quantity extraction expression fl to fn, rather than as a target based on the feature quantity as described above wherein the output value corresponding to the feature amount extraction expression [pi calculation expression to an amount of [pi fm [] to [] to determine the contribution rate the evaluation value fm. [0087] 例如,可以进行以下布置,其中通过将L条教导数据Ti的实际数据Di代入特征量提取表达式f 1而计算L个特征量f 1 [Di]之间的皮尔森相关系数,并且计算L条教导数据Ti的k类目标特征量Ili至iki;而且将计算的k个皮尔森相关系数的平均值确定为特征量提取表达式Π的评价值。 [0087] For example, arrangement may be made wherein the L teaching data by actual data Di Ti substitutes the feature amount extraction expressions f 1 is calculated Pearson correlation coefficient between the feature quantity of L f 1 [Di], and calculating the L teaching data Ti target feature amount k classes to Ili IKI; k and the average value calculated Pearson correlation coefficient is determined extraction expression evaluation value feature quantity of Π.

[0088] 另外,评价值计算单元14可以不仅计算相应特征量提取表达式Π至fm的评价值而且计算相应现有特征量Fl至Fm的评价值。 [0088] Further, the evaluation value calculating unit 14 can calculate not only the feature amount extraction expressions corresponding to the evaluation value Π and fm is calculated corresponding to the conventional feature amount evaluation value Fm, Fl.

[0089] 另外,评价值计算单元14基于从特征量计算单元12供应的相应特征量提取表达式f 1至fm的平均计算时间来校正这样确定的组成当前一代特征量提取表达式列表的相应特征量提取表达式Π至fm的评价值。 [0089] Further, the evaluation value calculating means 14 based on the respective characteristic amount calculating unit 12 is supplied from the feature amount extraction expressions f 1 to calculate an average correction time corresponding to fm features of such a composition to determine the current generation of feature amount extraction expression list amount extraction expression Π to evaluate the value of fm. 具体而言,评价值计算单元14将平均计算时间等于或者大于预定阈值的特征量提取表达式的评价值校正为其设置范围的最小值。 Specifically, the evaluation value calculating unit 14 calculates the average time is equal to or larger than a predetermined threshold value, the feature amount extraction expression evaluation value correction range set to its minimum. 随后,评价值计算单元14向特征量提取表达式列表生成单元11通知所校正的评价值。 Subsequently, the evaluation value calculating means 14 extracts the corrected evaluation value generation unit 11 to the notification list of expressions feature amount.

[0090] 根据这样的评价值校正,可以防止平均计算时间等于或者大于预定阈值的相关特征量提取表达式续传到下一代特征量提取表达式列表。 [0090] According to the evaluation value correction can be prevented average computation time is equal to or greater than a predetermined threshold value related to the feature amount extraction expressions forwarded to the feature amount extraction expression list of the next generation. 因而,随后可以减少下一代及其之后的特征量计算单元12的计算负荷。 Accordingly, the calculation load can be reduced then the feature quantity generation unit 12 after its calculation. 注意可以根据特征量计算单元12的计算能力来自动设置将与平均计算时间做比较的预定阈值,或者用户可以任意设置该预定阈值。 Note that calculation may be automatically set according to capacity of the feature quantity calculating unit 12 calculates the time will make the average of a predetermined threshold, or the user can arbitrarily set the predetermined threshold value.

[0091] 优化单元15容纳以下单元:特征量提取表达式优化单元16,用于优化从目标特征量计算表达式生成单元13提取的组成最后一代特征量提取表达式列表的特征量提取表达式fl至fm ;以及目标特征量计算表达式重构单元17,用于使用优化的特征量提取表达式fl至fm来重构目标特征量计算表达式。 [0091] The optimization unit 15 accommodates the following elements: feature amount extraction expression optimization unit 16 for optimization calculation expression consisting of the last generation of the feature amount generating unit 13 extracts the feature amount extraction expression lists of extracted feature amount from the target expression fl to fm; target feature quantity calculation expression and reconstruction unit 17, for using the optimized feature amount extraction expression fl to fm to reconstruct the target feature quantity calculation expression.

[0092] 特征量提取表达式优化单元16根据组成最后一代特征量提取表达式列表的相应特征量提取表达式fl至fm来检测预先登记的冗余算术运算的组合(下文称为“优化模式”),并且用处理负荷小的算术运算替换这些冗余算术运算,由此可以获得相同的算术运算结果,从而执行第一优化。 [0092] The feature amount extraction expression optimization unit 16 extracts a feature amount corresponding expression fl composition according to the last generation of the feature amount extraction expression list to a previously registered combination fm to detect redundant arithmetic operations (hereinafter referred to as "optimization mode ' ), and replaced with a small arithmetic operation processing load redundant arithmetic operation, thereby obtaining the same arithmetic operation result, thereby performing the first optimization. 下文将示出第一优化的例子。 The following shows an example of the first optimization.

[0093] 对于其中连续有计算绝对值的两个或者更多运算符Abs的优化模式,第二及其以后的运算符Abs是冗余的,因此通过用一个运算符Abs替换两个或者更多运算符Abs来执行优化。 [0093] wherein for continuous optimization mode with two or more operators calculated absolute value Abs, the second and subsequent operator Abs are redundant, and therefore an operator by using two or more alternative Abs Abs operator to perform optimization.

[0094] 对于其中连续有表示正规化算术运算的两个或者更多运算符Normalize的优化模式,第二及其以后的运算符Normalize是冗余的,因此通过用一个运算符Normalize替换两个或者更多运算符Normalize来执行优化。 [0094] wherein for two or more consecutive representing optimization mode operator Normalize normalized arithmetic operation, the second and subsequent operator Normalize are redundant, therefore replaced by two or by an operator Normalize Normalize more operators to perform the optimization.

[0095] 对于其中连续有表示平方运算的运算符Sqr和用于计算平方根的运算符Sqrt的优化模式,通过用处理负荷小的运算符Abs替换这两个运算符来执行优化,由此可以获得相同的算术运算结果。 [0095] wherein for continuous optimization mode indicating squaring operations and operators Sqr Sqrt operator for calculating a square root, the optimization is performed by replacing two operators with a small processing load Abs operator, thereby obtaining the the same arithmetic operation result.

[0096] 对于其中连续有表示微分运算的运算符Differential和表示积分运算的运算符Integrate的优化模式,通过消除运算符Differential和htegrate来执行优化,因为它们是不必要的。 [0096] For indicating where continuous operation of the differential operator Differential indicating optimization mode operator Integrate the integration operation is performed by removing the operators optimize Differential and htegrate, because they are unnecessary.

[0097] 注意优化模式及其优化方法不限于上述例子。 [0097] Note that the optimization model and optimization method is not limited to the above example.

[0098] 另外,特征量提取表达式优化单元16利用遗传算法来执行第二优化以便用更短的计算时间获得相同的计算结果。 [0098] Further, the feature amount extraction expression optimization unit 16 performs a second optimization in order to obtain the same results in a shorter computing time using genetic algorithms.

[0099] 目标特征量计算表达式重构单元17利用优化的特征量提取表达式Π至fm和教导数据通过机器学习来重构目标特征量计算表达式。 [0099] certain feature amount calculation unit 17 using the optimized expression reconstruction feature amount extraction expression Π to fm and the teaching data to reconstruct the target feature quantity calculation expression by machine learning.

[0100] 接着将参照图8中所示流程图对目标特征量计算表达式构造系统10的操作进行描述。 [0100] Next, the operation flowchart shown in FIG. 8 is configured in an expression system for a target computing feature amounts will be described with reference to FIG 10.

[0101] 在步骤Sl中,特征量提取表达式列表生成单元11随机生成组成第一代特征量提取表达式列表的m个特征量提取表达式,并且将由m个特征量提取表达式组成的特征量提取表达式列表供应到特征量计算单元12。 [0101] In step Sl, the feature amount extraction expression list generating unit 11 generates a random composition of the first generation feature quantity extraction expression list of the m feature amount extraction expression, and expression characteristics extracted by the feature quantity consisting of the m amount extraction expression lists supplied to the feature amount calculation unit 12.

[0102] 在步骤S2中,目标特征量计算表达式构造系统10获得教导数据Ti (i = 1、2直至L)。 [0102] In step S2, the target feature amount calculation expression system 10 is configured to obtain teaching data Ti (i = 1,2 until L). 所得教导数据Ti供应到目标特征量计算表达式生成单元13和优化单元15。 The resulting Ti teachings data supplied to the target feature amount calculation unit 13 and the optimizing expression generating unit 15.

[0103] 在步骤S3中,特征量计算单元12将教导数据Ti中包括的实际数据Di代入组成从特征量提取表达式列表生成单元11供应的特征量提取表达式列表的相应特征量提取表达式f 1至fm以计算(LXm)个特征量fj [Ti],也测量相应特征量提取表达式f 1至fm的计算所必需的计算时间,并且在通过将不同的L条实际数据Di代入每个特征量提取表达式来执行计算时计算平均计算时间。 [0103] In step S3, the feature amount calculation unit 12 on behalf of the teachings of the actual data Di included in the data consisting of Ti extraction expression list generating unit 11 supplies the feature quantity from the feature amount extraction expression lists of the respective feature amount extraction expressions f 1 to fm in feature quantity calculation (LXm) fj [Ti], also measured corresponding feature amount extraction expression calculation time f calculate 1 to fm necessary, and by different the L actual data Di is substituted into each of the a feature amount extraction expressions is performed when calculating the average time is calculated. 算出的与特征量提取表达式Π至fm各自对应的(LXm)个特征量fj [Ti]和平均计算时间供应到目标特征量计算表达式生成单元13。 Calculated feature amount extraction expressions each corresponding to Π to fm (LXM) feature amounts fj [Ti] and the average computation time is supplied to the target feature quantity calculation expression generating unit 13.

[0104] 在步骤S4中,目标特征量计算表达式生成单元13通过利用特征选择的线性鉴别或者递归、根据作为特征量计算单元12的计算结果的(LXm)个特征量fj [Ti]来学习用于估计教导数据Ti中包括的L个目标特征量Ili的目标特征量计算表达式。 [0104] In step S4, the target characteristic quantity calculation expression generating unit 13 by using the selected feature or a recursive linear discriminant, based on the calculation result of the calculating unit 12 (LXM) feature amount as the feature quantity fj [Ti] to learn teaching data for estimating the L Ti included in the target feature amount Ili target feature amount calculation expression.

[0105] 现在将参照图9中所示流程图对目标特征量计算表达式生成单元13在步骤S4中的处理(下文称为学习处理)具体地进行描述。 [0105] Expression of the target feature amount calculation unit 13 generates the processing in step S4 (hereinafter referred to as a learning process) will now be specifically described with reference to the flowchart shown in FIG.

[0106] 在步骤S21中,在生成目标特征量计算表达式时,目标特征量计算表达式生成单元13随机生成多个选择表TB,这些选择表TB表示η个现有特征量Fl至1¾和作为m个特征量提取表达式fl至fm的输出的特征量fl □至fm[]之中的已用(被选)特征量和未用(未选)特征量,由此生成第一代选择表组。 [0106] In step S21, when the target generation feature quantity calculation expression, the target characteristic quantity calculation expression generating unit 13 generates a plurality of random selection table TB, the selection table TB indicates a conventional feature amount η 1¾ to Fl and as the m feature amount extraction expressions fl fm to output the feature amount has been used to fl □ (selected) feature amounts and without (non-selected) feature amounts, thereby generating a first generation of [] is selected from among the fm table group. 将组成选择表组的多个选择表TB作为基因,基于遗传算法在下文描述的步骤S29中更新这一选择表组。 The composition of selecting a plurality of selected groups of Table TB gene, update the selection table set in step genetic algorithm based on S29, described below.

[0107] 图10图示了将要生成的由多个选择表TB组成的选择表组的例子。 Examples of the group selection table selected by the table TB composed of a plurality of [0107] FIG. 10 illustrates a to be generated. 注意在图10 中圆标记表示被选的而X标记表示未选的。 Note that the circle mark in FIG. 10 shows the selected and unselected X numerals.

[0108] 在步骤S22中,目标特征量计算表达式生成单元13在关注组成当前一代选择表组的相应选择表TB之时一次一个依次开始选择表组循环。 [0108] In step S22, the target characteristic quantity calculation expression generating unit 13 sequentially one at a time in the composition of interest began to select the current generation cycle table group selection table to select the appropriate set of tables TB. 注意选择表组循环的重复次数为组成选择表组的选择表TB的数目(图10所示例子中的χ)。 Note that the group selection list is the number of repetitions of the cycle consisting of the selection list of the group selection table TB ([chi] in the example shown in FIG. 10).

[0109] 在步骤S23中,目标特征量计算表达式生成单元13利用从特征量计算单元12供应的与特征量提取表达式Π至fm各自对应的平均计算时间来确定与关注的选择表TB所选择的特征量fj[]对应的特征量提取表达式fj的平均计算时间的总和是否等于或者小于预定阈值。 [0109] In step S23, the target characteristic quantity calculation unit 13 calculates expression generating unit 12 and the feature quantity supplied from the feature amount extraction expressions to the average computation time fm Π each corresponding to the determined table TB is selected. the selected feature amount fj [] if the sum of the average computation time corresponding to the feature amount extraction expression fj is equal to or less than a predetermined threshold value. 可以根据目标特征量计算表达式生成单元13的计算能力来自动设置将要与平均计算时间的总和做比较的预定阈值,或者用户可以任意设置该预定阈值。 Computing capacity can expression generating unit 13 calculates a target characteristic quantity to be set automatically according to a predetermined threshold value is compared with the sum of the average calculation time, or the user can arbitrarily set the predetermined threshold value.

[0110] 在确定平均计算时间的总和等于或者小于预定阈值的情况下,该处理继续步骤S24。 [0110] In the case of determining the average computation time is equal to or less than a sum of a predetermined threshold value, the process proceeds to step S24.

[0111] 在步骤SM中,目标特征量计算表达式生成单元13利用在作为特征量计算单元12 的计算结果的(LXm)个特征量fj [Ti]和在教导数据Ti中包括的(LXn)个现有特征量Fli 至I^i之中由关注的选择表TB选择的特征量、通过线性鉴别或者递归来学习数目与目标特征量类型的数目(k)相等的目标特征量计算表达式。 [0111] In step SM, the target characteristic quantity calculation expression generating unit 13 uses the calculation result fj unit 12 calculates a feature quantity of the feature quantity (LXM) [Ti] and the teaching data included Ti (LXN) among the two prior to the feature amount Fli I ^ i selected by the feature quantity selection table TB of interest to learn the number (k) of the number of feature quantity type or a recursive object is achieved by a linear discriminant equal to target feature amount calculation expression.

[0112] 在步骤S25中,目标特征量计算表达式生成单元13将步骤SM中处理的学习结果的Akaike信息标准(AIC)计算为所关注的选择表TB的评价值。 [0112] In step S25, the evaluation value feature quantity calculation target of interest for the selection table TB calculated Akaike Information Standards (AIC) expression learning result generated in step 13 in the processing unit SM.

[0113] 注意在步骤S23中确定与关注的选择表Tb所选择的特征量fj[]对应的特征量提取表达式fj的平均计算时间的合计大于预定阈值的情况下,该处理继续步骤S26。 In the case [0113] Note that the characteristic amount determining of interest in step S23 selection table Tb selected fj [] corresponding to the feature amount extraction average computation time expression fj is the sum is greater than a predetermined threshold value, the process proceeds to step S26. 在步骤S^中,目标特征量计算表达式生成单元13将关注的选择表TB的评价值设置为其设置范围的最小值。 In step S ^, the target minimum evaluation value feature quantity calculating unit 13 is provided to generate an expression of interest selected table TB to set the range. 因此,防止平均计算时间的合计大于预定阈值的选择表续传到下一代,由此可以防止延长用于计算将要生成的目标特征量计算表达式所必需的时间。 Accordingly, to prevent the average computation time is greater than the sum of a predetermined threshold value selection table forwarded to the next generation, thereby calculating the target characteristic quantity to be generated for the extension of the time necessary for the calculation expression may be prevented.

[0114] 在通过步骤S25或者步骤幻6中的处理来确定所关注的选择表TB的评价值之后, 该处理继续步骤S27。 [0114] After the process in step 6 S25 or the step of determining by phantom evaluation value selection table TB of interest, the process proceeds to step S27. 在步骤S27中,目标特征量计算表达式生成单元13确定已经关注组成当前一代选择表组的所有选择表TB,而在有尚未关注的选择表TB情况下,该处理返回到步骤S22,由此重复步骤S22至S27中的处理。 In step S27, the target characteristic quantity calculation expression generating unit 13 determines the composition of the current generation has been of interest for all the group selection table selection table TB, and in the case where there is not yet selected the table TB of interest, the process returns to step S22, the thus repeat the processing in steps S22 to S27. 随后在步骤S27中,在已经关注组成当前一代选择表组的所有选择表TB情况下,该处理继续步骤S28。 Then in step S27, the selection table has been concerned about all TB cases that comprise the current generation of group selection table, the process proceeds to step S28.

[0115] 在步骤S^中,目标特征量计算表达式生成单元13确定是否已经为预定若干代改进了评价最充分的选择表TB的评价值。 [0115] In step S ^, the target characteristic quantity calculation expression generating unit 13 determines whether the evaluation value for evaluating the most improved sufficiently to a predetermined selection table TB several generations. 随后在确定已经改进了评价最充分的选择表TB的评价值情况下,或者在确定自从评价值的改进停止起尚未过去预定若干代的情况下,该处理继续步骤S29。 Subsequently determined to have been improved in a case the evaluation value evaluated fullest selection table TB, or improved since the evaluation value is determined to stop the situation from several generations of the predetermined has not passed, the processing proceeds to step S29.

[0116] 在步骤S29中,目标特征量计算表达式生成单元13通过使用基于每个选择表TB 的评价值的遗传算法更新当前一代选择表组来生成下一代选择表组。 [0116] In step S29, the target characteristic quantity calculation expression generating unit 13 updates the current generation of the selection list to generate the next set of genetic algorithm selection table group evaluation value based on each selected by use of the table TB. 该处理返回到步骤S22,由此重复后续处理。 The process returns to step S22, the subsequent processing is repeated thereby.

[0117] 随后在步骤幻8中,在确定尚未针对预定若干代改进评价最充分的选择表TB的评价值情况下,该处理继续图8中所示步骤S5。 [0117] Then, in Step 8 in phantom, has not been improved in determining the most adequate choice evaluation table TB for several generations predetermined evaluation value, the process continues as shown in FIG. 8 step S5.

[0118] 根据上述学习处理,已经生成用于计算与当前一代特征量提取表达式列表对应的k类目标特征量中每一类目标特征量的目标特征量计算表达式。 [0118] According to the above-described learning process, it has been generated for calculating the target current generation feature quantity extraction target feature amount based expression list feature amount k corresponding to each type of the target feature quantity calculation expression.

[0119] 注意对于上文提到的描述,已经在遗传搜索方法和ACI用于学习处理的假设下进行了描述,但是可以利用不同方法来执行学习处理。 [0119] Note that for the above-mentioned description, it has been assumed for the learning process are described in the genetic search method and the ACI, but using different methods to perform a learning process. 另外,可以利用局部搜索而不是遗传算法来确定现有特征量的选择或者未选或者特征量提取表达式的输出值。 Further, instead of using a local search algorithm to determine the existing genetic feature amount selected or unselected feature amount extraction expression or output value.

[0120] 例如,在利用局部搜索的情况下,在所有的η个现有特征量Fl至1¾和作为m个特征量提取表达式fl至fm的输出的特征量fl[]至fm[]未选的情况下开始学习。 [0120] For example, in the case of using a local search, all of the feature amount in a conventional η 1¾ to Fl feature amount extraction expression and the m feature amounts as fl to fm outputted fl [] to fm [] No He began studying the case of the election. 随后在η个现有特征量Fl至1¾和作为m个特征量提取表达式Π至fm的输出的特征量Π []至fm[]中一个特征量被选择而其它未选的情况下生成(n+m)个选择表,并且使用AIC对每个选择表执行评价。 Followed by a conventional η 1¾ to Fl feature amount and a feature amount extraction expressions of m [pi fm to output the feature quantity [pi [] is selected to fm [] a feature amount generating otherwise unselected ( n + m) th table selected, and performing evaluation of each AIC selection table to use. 随后确定评价最充分、即ACI值小的选择表。 Then determine the evaluation fullest, namely small ACI value selection table. 另外,利用η个现有特征量Fl至1¾和作为m个特征量提取表达式f 1至fm的输出的特征量f 1 []至fm[]之中的一个特征量被选择而其它未被选的情况下生成(n+m)个选择表,并且使用AIC等对每个选择表执行评价。 Further, by using a conventional η 1¾ to Fl and feature amount as the feature amount extraction expressions of m f 1 fm to output the feature amount f 1 [] to fm [] is a feature quantity selected from among the other not generating a case where the selected (n + m) th selection table, and the like using the AIC performed evaluation of each selection list. 优选地重复上述处理直至AIC等评价的改进停止。 Improvement of the above process is preferably repeated until the evaluation is stopped and the like AIC.

[0121] 现在将回到对图8的描述。 [0121] Now description will be back to FIG. 8. 在步骤S5中,目标特征量计算表达式生成单元13的评价值计算单元13计算当前生成的k个相应目标特征量计算表达式的作为相应特征量提取表达式Π至fm的计算结果的特征量Π []至fm[]的贡献率,并且将通过合计所计算的k个贡献率而得到的总贡献率确定为组成当前一代特征量提取表达式列表的相应特征量提取表达式fl至fm的评价值。 In step S5, the target characteristic quantity calculation expression generating unit 13, an evaluation value calculating unit 13 calculates current corresponding to the k target generated feature quantity calculation expression as respective feature amount extraction expression Π to the calculation result of the feature quantity fm [pi [] to fm [] contribution rate, and the sum obtained by the calculated contribution ratio of k determines the total contribution rate corresponding to the composition of the current generation of feature amount extraction expression list feature amount extraction expression fl to fm, Evaluation value.

[0122] 注意在步骤S5中,可以基于皮尔森相关系数而不是如上所述基于目标特征量计算表达式的作为相应特征量提取表达式Π至fm的输出的特征量Π □至fm[]的贡献率来确定组成当前一代特征量提取表达式列表的相应特征量提取表达式Π至fm的评价值。 [0122] Note that in step S5, based on the Pearson correlation coefficient calculation expression as described above rather than as respective outputs feature amount extraction expression to [pi fm based on the target feature quantities Π □ to fm [] of the contribution rate to determine the composition of the evaluation value corresponding feature amounts of the current generation feature quantity extraction expression lists of extraction expression Π to fm.

[0123] 在步骤S6中,评价值计算单元14基于从特征量计算单元12供应的相应特征量提取表达式fl至fm的平均计算时间来校正在步骤S5中的处理中确定的组成当前一代特征量提取表达式列表的相应特征量提取表达式f 1至fm的评价值。 [0123] In step S6, the evaluation value calculating means 14 based on the respective feature amount calculation unit 12 supplies the feature amount extraction expressions from fl to fm average computation time is corrected in step S5 in the processing composition determined current generation feature amount extraction expression list corresponding to the evaluation feature amount extraction expressions f 1 to a value of fm. 具体而言,评价值计算单元14将平均计算时间等于或者大于预定阈值的特征量提取表达式的评价值校正为其设置范围的最小值。 Specifically, the evaluation value calculating unit 14 calculates the average time is equal to or greater than a predetermined threshold value, the feature amount extraction expression evaluation value correction range set to its minimum. 随后,评价值计算单元14向特征量提取表达式列表生成单元11通知所校正的评价值。 Subsequently, the evaluation value calculating means 14 extracts the corrected evaluation value generation unit 11 to the notification list of expressions feature amount.

[0124] 在步骤S7中,目标特征量计算表达式生成单元13确定当前生成的目标特征量计算表达式的计算结果是否已经达到所需精确度或者用户是否已经执行结束操作。 [0124] In step S7, the target characteristic quantity calculation expression generating unit 13 determines a target generated current calculation feature if the expression evaluates whether the desired accuracy has been reached or the user has performed an end operation. 在确定计算结果尚未达到所需精确度而用户也尚未执行结束操作的情况下,该处理继续步骤S8。 In determining the calculation result has not reached the desired accuracy and the user end operation has not been performed, the processing proceeds to step S8.

[0125] 在步骤S8中,特征量提取表达式列表生成单元11通过根据遗传算法更新当前一代特征量提取表达式列表来生成下一代特征量提取表达式列表。 [0125] In step S8, the feature amount extraction expression list of the next generation unit 11 generates the feature amount extraction expression list of the current generation by the feature amount extraction expression lists based on the genetic algorithm update. 随后,该处理返回到步骤S3,其中重复步骤S3及其之后步骤中的处理。 Subsequently, the process returns to step S3, wherein steps S3 and the processing after step.

[0126] 随后,在步骤S7中确定来自当前生成的目标特征量计算表达式的计算结果已经达到所需精确度或者用户已经执行结束操作的情况下,该处理继续步骤S9。 [0126] Subsequently, in step S7 is determined from the current target feature amount calculation result of the expression generated has reached the required accuracy or the case where the user has performed the ending operation, the process proceeds to step S9.

[0127] 在步骤S9中,目标特征量计算表达式生成单元13将当前一代特征量提取表达式列表和当前生成的目标特征量计算表达式作为最后一代特征量提取表达式列表和与之对应的k个目标特征量计算表达式输出到优化单元15。 [0127] In step S9, the target characteristic quantity calculation expression generating unit 13 of the current generation of feature amount extraction expression list and a target generated current calculation expression feature amount extraction expression list of the final generation as a feature quantity and the corresponding k target feature quantity calculation expression of the output to the optimization unit 15.

[0128] 在步骤SlO中,优化单元15优化从目标特征量计算表达式生成单元13输入的最后一代特征量提取表达式列表的相应特征量提取表达式Π至fm、并且也利用所优化的相应特征量提取表达式Π至fm来重构目标特征量计算表达式。 [0128] In step SlO, the optimization of the optimization unit 15 calculates the final generation feature quantity generation unit 13 expression extraction expression lists inputted feature amount extraction expressions corresponding to Π FM, and is also optimized using the corresponding feature quantity from the target Π feature amount extraction expression to be reconstructed fm target feature quantity calculation expression.

[0129] 现在将参照图11中所示流程图对优化单元15在步骤SlO中的处理具体地进行描述。 [0129] Referring now to the flowchart shown in FIG. 11 to be described specifically optimization unit 15 in the processing in step SlO.

[0130] 在步骤S41中,特征量提取表达式优化单元16根据组成最后一代特征量提取表达式列表的相应特征量提取表达式π至fm来检测优化模式,并且执行以下第一优化,该第一优化用于用处理负荷小的算术运算替换处理负荷大的算术运算,由此可以获得相同的算术 [0130] In step S41, the feature amount extraction unit 16 extracts optimized expression to expression π fm optimization mode is detected according to the corresponding feature quantity consisting of the last generation of the feature amount extraction expression lists, and following a first optimization, the second optimization processing for replacing a large load arithmetic operation with a small load of arithmetic processing, thereby obtaining the same arithmetic

运算结果。 Result of the operation.

[0131] 在步骤S42中,特征量提取表达式优化单元16在第一优化之后将遗传算法应用于相应特征量提取表达式fl至fm以便以更短的计算时间获得相同的计算结果,由此执行第二优化。 [0131] In step S42, the feature amount extraction expression optimizing unit 16 after the first optimization algorithm is applied to the appropriate genetic characteristic amount extraction expressions fl to fm in order to obtain the same results in a shorter computing time, thereby performing a second optimization.

[0132] 现在将参照图12中所示流程图对特征量提取表达式优化单元16在步骤S42中的处理具体地进行描述。 [0132] Optimization of expression extraction unit 16 the feature quantity processing in step S42 will now be described specifically with reference to the flowchart shown in FIG. 12.

[0133] 在步骤S51中,特征量提取表达式优化单元16在第一优化之后在关注组成最后一代特征量提取表达式列表的相应特征量提取表达式Π至fm之时一次一个依次开始特征量提取表达式列表循环。 [0133] In step S51, the feature amount extraction expression optimization unit 16 in the last generation after the first optimization feature amount extraction expression list corresponding to the feature amount extraction expression Π fm when to start sequentially one at a time feature amount of the attention composition extraction expression list loop. 注意特征量提取表达式列表循环的重复次数为组成特征量提取表达式列表的特征量提取表达式Π至fm的数目m。 Note that the number of repetitions feature amount extraction expression list loop is composed of the feature amount extraction expression list to the feature amount extraction expression Π fm, the number m. [0134] 在步骤S52中,特征量提取表达式优化单元16变异所关注的特征量提取表达式fj 的一部分以生成R个优化候选表达式f = 1、2直至R)、并且将这些表达式确定为第一代优化候选表达式组。 [0134] In step S52, the feature amount extraction expression optimizing unit 16 variation of interest feature amount extraction expression fj part candidate optimization to generate R f = 1,2 until expression R), and these expressions The first generation is determined as candidates for optimization expression set.

[0135] 在步骤S53中,特征量提取表达式优化单元16将S条评价数据(该数据具有与输入数据C相同的类型)代入所关注的特征量提取表达式fj以计算S个特征量fj[]。 [0135] In step S53, the feature amount extraction expression optimizing unit 16 reviews the S data (data having the same type of the input data C) into the features of interest fj amount extraction expression to calculate a feature quantity fj S [].

[0136] 在步骤SM中,特征量提取表达式优化单元16在关注组成当前一代优化候选表达式组的R个优化候选表达式之时一次一个依次开始优化候选表达式组循环。 [0136] In step SM, the feature amount extraction unit 16 to optimize expression of interest to optimize the composition of the current generation of R sequentially one at a time starting the optimization cycle when optimizing candidate expression set of candidate expressions candidate expression group. 注意优化候选表达式组循环的重复次数为组成优化候选表达式组的优化候选表达式的数目。 Note that optimize expression set repetitions candidate optimization cycle was composed of the number of candidates for optimization of expression set of candidate expressions.

[0137] 在步骤S55中,特征量提取表达式优化单元16将步骤S53中利用的S条评价数据代入所关注的优化候选表达式fj' r以计算S个特征量fj' J]、也在代入相应多条评价数据时测量计算时间、并且还计算平均计算时间。 [0137] In step S55, the feature amount extraction expression optimizing candidate expression fj optimization unit 16 in step S53 using the evaluation data into pieces S of interest 'r to compute a feature amount S fj' J], also calculated measurement time is substituted into the corresponding plurality of evaluation data, and also calculates the average computation time.

[0138] 在步骤S56中,特征量提取表达式优化单元16计算表示了作为步骤S53中处理结果的S个特征量fj[]与作为步骤S55中处理结果的S个特征量fj'J]之间相关度的皮尔森相关系数、并且确定S个特征量fj[]与S个特征量fj' J]之间相关度是否近似为1.0。 [0138] In step S56, the feature amount extraction expression optimizing unit 16 calculates a processing in step S53 as a result of the S feature amounts FJ [] and as a result of the processing step S55 of the S feature amounts fj'J] of Pearson correlation between the correlation coefficient and determines the S feature amounts fj [] and the S feature amount of correlation between fj 'J] are approximately 1.0. 随后,在确定S个特征量fj []与S个特征量fj\[]之间相关度近似为1. 0的情况下,该处理继续步骤S57。 Subsequently, the feature amount is determined the S FJ [] and the S feature amounts FJ \ [correlation between] 1.0 is approximately the case, the process continues with step S57.

[0139] 在步骤S57中,特征量提取表达式优化单元16将步骤S55的处理中计算的平均计算时间的倒数设置为关注的优化候选表达式幻'的评价值。 [0139] In step S57, the feature amount extraction expression optimizing unit 16 will set the evaluation value of the inverse of the step candidate optimization expression magic 'is a time average calculation of interest calculated in the processing S55.

[0140] 注意在步骤S56中确定S个特征量fj[]与S个特征量fj'J]之间相关度是否近似为1.0。 [0140] Note that determines a feature amount S FJ [] and the correlation between the S feature amounts fj'J] are approximately 1.0 in step S56. 随后,在确定S个特征量fj[]与S个特征量fj'J]之间相关度并不近似为1.0 的情况下,该处理继续步骤S58。 Subsequently, the feature amount is determined between the S FJ [] and the S feature amounts fj'J] Relevance not be approximated to 1.0, the process proceeds to step S58.

[0141] 在步骤S58中,特征量提取表达式优化单元16将关注的优化候选表达式fj'的评价值设置为其范围的最小值。 Evaluation value [0141] In step S58, the feature amount extraction expression optimizing unit 16 will focus on optimizing the candidate expression fj 'for which the minimum range.

[0142] 在通过步骤S57或者步骤S58中的处理来确定所关注的优化候选表达式fj' r的评价值之后,该处理继续步骤S69。 After [0142] In the processing in step S58 through step S57 to determine or optimize candidate interest fj 'expression evaluation value r, the process proceeds to step S69. 在步骤S59中,特征量提取表达式优化单元16确定是否已经关注组成当前一代优化候选表达式组的所有优化候选表达式fj'r,而在有尚未关注的优化候选表达式fj' r的情况下,该处理返回到步骤S54,其中重复步骤SM至S59中的处理。 In step S59, the feature amount extraction expression optimizing unit 16 determines whether the current generation of optimized composition of interest for all candidates for optimization of expression fj'r candidate expression set, while there is not yet optimized candidate expressions of interest fj 'r, where , the processing returns to step S54, wherein the repeating step S59 to the processing in the SM. 随后在步骤S69中,在已经关注组成当前一代优化候选表达式组的所有优化候选表达式fj' r的情况下,该处理继续步骤S60。 Then, in step S69, the current generation has been of interest to optimize the composition of all candidate expression set candidates for optimization expression fj 'in the case of r, the process proceeds to step S60.

[0143] 在步骤S60中,特征量提取表达式优化单元16确定是否已经为预定若干代改进了评价最充分的优化候选表达式的评价值。 [0143] In step S60, the feature amount extraction unit 16 determines whether the expression has been optimized to improve the fullest expression of the optimal evaluation value candidate was evaluated as a number of predetermined generations. 随后,在确定已经改进了评价最充分的优化候选表达式的评价值的情况下,或者在确定自从评价值的改进停止起尚未过去预定若干代的情况下,该处理继续步骤S61。 Subsequently, in determining evaluation has improved the most fully optimized case expression evaluation value of the candidate, or in determining the value of the evaluation improved since the stop of the case has not been scheduled in the past several generations, the processing continues to step S61.

[0144] 在步骤S61中,特征量提取表达式优化单元16通过使用基于每个优化候选表达式f j' r的评价值的遗传算法更新当前一代优化候选表达式组来生成下一代优化候选表达式组。 [0144] In step S61, the feature amount extraction expression optimizing unit 16 by using the expression set of candidates for optimization Genetic Algorithm optimization for each 'candidate expression evaluation value r fj update the current generation to generate the next candidate optimize expression group. 然而注意进行以下布置,其中在下一代优化候选表达式组中包括评价最充分的优化候选表达式,而通过变异具有最充分评价的前一代优化候选表达式和关注的特征量提取表达式fj各自的一部分来生成区域(RI)个优化候选表达式。 Note however that the following are arranged, wherein in optimizing the next group of candidate expressions including the most adequate candidate optimal evaluation expression, and having a feature amount variation by optimizing the candidate predecessor fullest expression of interest and the evaluation of their extraction expression fj generating a region portion (RI) a candidate optimization expression. 该处理返回到步骤S54以重复后续处理。 The process returns to step S54 to repeat the subsequent processing. 1[0145] 随后,在步骤S60中确定尚未针对预定若干代改进评价最充分的优化候选表达式的情况下,该处理继续步骤S62。 1 [0145] Subsequently, it is determined yet improved for several generations predetermined optimal evaluation fullest expression of the candidate in the case where the step S60, the process proceeds to step S62.

[0146] 根据步骤S52至S60中的处理,在第一优化之后组成最后一代特征量提取表达式列表的所有特征量提取表达式之一已经受到第二优化。 [0146] The steps S52 through S60 in the processing, optimization of the composition after all of the first feature quantity of the last generation of the feature amount extraction expression list to extract a second one has been optimized expression.

[0147] 在步骤S62中,特征量提取表达式优化单元16确定在第一优化之后是否已经关注组成最后一代特征量提取表达式列表的所有特征量提取表达式Π至fm,而在有尚未关注的特征量提取表达式fj的情况下,该处理返回到步骤S51,其中重复步骤S51至S62中的处理。 [0147] In step S62, the feature amount extraction expression optimizing unit 16 determines whether after the first optimization composition of all concern the last generation feature quantity extraction expression list feature amount extraction expression Π to fm, have not been of interest in in the case of the feature amount extraction expression fj, the process returns to step S51, the processing in S51 to S62 wherein steps. 随后在步骤S62中,在第一优化之后已经关注了组成最后一代特征量提取表达式列表的所有特征量提取表达式Π至fm的情况下,这意味着在第一优化之后组成最后一代特征量提取表达式列表的所有特征量提取表达式fl至fm已经受到第二优化,因此该处理继续图11中所示步骤S43。 In the case of subsequent step S62, after the first concern has been to optimize the amount of the composition of all the features of the last generation feature quantity extraction expression list of extraction expression Π to fm, which means that the composition of the last generation of feature amount after the first optimization All feature amount extraction expression list extraction expression fl to fm second has been optimized, so in step 11 the process continues as shown in FIG S43.

[0148] 现在将回到对图11的描述。 [0148] Description will now return to FIG. 11. 在步骤S43中,目标特征量计算表达式重构单元17基于受到第二优化的特征量提取表达式Π至fm以及基于教导数据、通过统计分析和机器学习来重构从目标特征量计算表达式生成单元13供应的k个目标特征量计算表达式。 In step S43, the target characteristic quantity calculation expression data reconstruction unit 17 extracts the teachings expression Π to fm and based on the second feature amount being optimized to reconstruct through statistical analysis and machine learning target feature amount is calculated from the expression generating unit 13 supplies the k target feature amount calculation expression. 随后, 输出最终的特征量提取表达式列表和目标特征量计算表达式,并且结束目标特征量计算表达式构造系统10的操作。 Subsequently, a final output characteristic amount extraction expression list of the target and the feature amount calculation expression, and terminates the operation target feature amount calculation expression system 10 configuration.

[0149] 对目标特征量计算表达式构造系统10的操作的说明到此为止。 [0149] The system configuration described calculation expression of the target feature amount of the operation 10 so far.

[0150] 如上所述,根据已经应用本发明一个实施例的目标特征量计算表达式构造系统10,使用现有特征量来自动构造目标特征量计算表达式,由此与仅利用特征量计算表达式的计算结果来构造目标特征量计算表达式的情况相比可以减少往往变得冗余的特征量提取表达式的数目。 [0150] As described above, the computing system configured in accordance with the expression amount of the target feature has been applied to one embodiment of the present invention 10, the conventional feature amount calculated using the target feature amount to automatically construct an expression, whereby the expression of only using the feature amount calculation formula calculation result of the target feature quantity calculation expression construct compared to a case often can reduce the number of feature amount extraction expression becomes redundant. 因此在根据目标特征量计算表达式来计算目标特征量时可以减少算术运算的处理量。 Therefore, when the target feature quantity is calculated based on the target feature amount calculation expression arithmetic operation processing amount can be reduced.

[0151] 另外,根据已经应用本发明一个实施例的目标特征量计算表达式构造系统10,通过遗传算法来防止计算时间久的特征量计算表达式的基因续传到后代,由此可以构造算术运算时间受限制的目标特征量计算表达式。 Gene [0151] In addition, the configuration of the system according to the calculation expression has been the target feature amount embodiment applying the present invention is an embodiment 10, the genetic algorithm to prevent long calculation time of the feature amount calculation expression forwarded to progeny, whereby texture arithmetic limited operation time target feature amount calculation expression. 也可以缩短目标特征量计算表达式的构造时间。 Calculating construction time can be shortened expression of the target feature quantity.

[0152] 另外,根据已经应用本发明一个实施例的目标特征量计算表达式构造系统10,优化了组成特征量提取表达式列表的特征量提取表达式,由此在根据目标特征量计算表达式来计算目标特征量时可以减少算术运算的处理量。 [0152] Further, according to the calculation expression system configured target feature amount has been applied to one embodiment of the present invention 10, to optimize the composition of the feature amount extraction expression lists of the characteristic amount extraction expression, whereby the feature quantity calculation expression according to the target arithmetic processing amount can be reduced when the target feature amount calculated.

[0153] 附带提一点,上述连串处理可以通过硬件实现、也可以通过软件实现。 [0153] Incidentally, the above-described series of processing may be implemented by hardware, may be realized by software. 在通过软件实现连串处理的情况下,组成其软件的程序从程序记录介质安装到例如能够通过安装各类程序来执行各类功能的专用硬件中嵌入的计算机或者通用个人计算机等。 In case of realizing the series of processing by software, a program which is composed of software installed from a program recording medium, for example, through installation of various programs to execute various kinds of functions in dedicated hardware computer or embedded general-purpose personal computer or the like.

[0154] 图13是使用程序来执行上述连串处理的计算机的硬件配置例子的框图。 [0154] FIG. 13 is a block diagram of using a program to execute the above-described series of processing hardware configuration example of a computer.

[0155] 利用这一计算机100,CPU(中央处理单元)101、ROM(只读存储器)102和RAM(随机存取存储器)103通过总线104相互连接。 [0155] With this computer 100, CPU (Central Processing Unit) 101, ROM (Read Only Memory) 102 and RAM (Random Access Memory) 103 are interconnected by a bus 104.

[0156] 输入/输出接口105还连接到总线104。 [0156] Input / output interface 105 is also connected to the bus 104. 输入/输出接口105与包括键盘、鼠标、 麦克风等的输入单元106、包括显示器、扬声器等的输出单元107、包括硬盘、非易失性存储器等的存储单元108、包括网络接口的通信单元109和驱动诸如磁盘、光盘、磁光盘、半导体存储器等可拆卸记录介质111的驱动器110连接。 The input / output interface 105 includes a keyboard, a mouse, a microphone, etc. The input unit 106 including a display, a speaker, an output unit 107 including a hard disk, a nonvolatile memory such as the storage unit 108 including a network interface and a communication unit 109 drives a removable recording medium such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory medium 111 of the drive 110 are connected.

20[0157] 利用这样配置的计算机,CPUlOl通过输入/输出接口105和总线104在RAM103中加载例如存储器单元108中存储的程序来执行该程序,由此执行上述连串处理。 20 [0157] using the computer thus configured, CPUlOl e.g. loading a program stored in the memory unit 108 in the RAM103 through the input / output interface 105 and the bus 104 to execute the program, thereby performing the above series of processing.

[0158] 注意计算机所执行的程序可以是其中按照在本说明书中描述的顺序以时间顺序执行处理的程序或者可以是其中并行地或者按照诸如在被调用时等必需时序执行处理的程序。 Programs [0158] Note that may be performed by the computer executing a program wherein the processing in time series in the order described in the present specification or may be parallel or which executes processing according to the timing and the like as required when the program is called.

[0159] 另外,该程序也可以由单个计算机处理或者可以由多个计算机以分布方式处理。 [0159] Further, the program may be processed by a single computer or may be processed in a distributed manner by plural computers. 另外,该程序可以传送到远程计算机进行执行。 Further, the program may be transferred to a remote computer for execution.

[0160] 另外在本说明书中,术语“系统”代表由多个设备配置而成的整个设备。 [0160] Further, in the present specification, the term "system" represents the device configured by a plurality of the entire apparatus formed.

[0161] 注意本发明的实施例不限于上述实施例,并且可以执行各种修改而不脱离本发明的实质。 [0161] Note that the present embodiment of the invention is not limited to the above embodiments, and various modifications may be performed without departing from the spirit of the invention.

[0162] 本领域技术人员应当理解根据目前设计要求和其它因素可以想到在所附权利要求及其等效的范围内的各种修改、组合、二次组合和变更。 [0162] It should be understood by those skilled in the art based on the current design requirements and other factor may occur in the appended claims and their various modifications, combinations, sub-combinations and alterations within the equivalent range.

Claims (13)

1. 一种信息处理设备,用于获取输入数据和与所述输入数据对应的现有特征量作为输入、并且生成用于输出与所述输入数据对应的目标特征量的目标特征量计算表达式,所述信息处理设备包括:特征量提取表达式列表生成装置,配置成将在前一代的包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于所述特征量提取表达式的评价值的遗传算法更新前一代的所述特征量提取表达式列表来生成所述特征量提取表达式列表;特征量计算装置,配置成将作为教导数据供应的实际数据输入到在所述特征量提取表达式列表中包括的每个特征量提取表达式以计算与所述实际数据对应的多个特征量;目标特征量计算表达式生成装置,配置成同等地利用所计算的与所述实际数据对应的所述多 1. An information processing apparatus for acquiring the input data and the input data corresponding to the conventional feature amount as an input and generates a feature amount of the target output and the input data corresponding to the target feature amount calculation expression for said information processing apparatus comprising: feature amount extraction expression list generation means is configured to extract a plurality of feature amounts include a plurality of feature amount extraction expression operators composed of first generation expression list comprises a plurality feature quantity extraction for gene expression, extraction expression evaluation value by using the feature amount based on the genetic algorithm to update the feature amount extraction expression lists of the previous generation to generate the feature amount extraction expression list; feature amount computing means configured to input the actual data to the teaching data supplied from the feature amount extraction expression lists each including extraction expression to calculate the actual data corresponding to the plurality of feature amounts in the feature amount; target feature the expression generation amount calculation device configured to equally utilize the calculated and the actual data corresponding to the plurality 特征量和与作为教导数据供应的所述实际数据对应的现有特征量,通过用于估计与作为教导数据供应的所述实际数据对应的目标特征量的机器学习来生成所述目标特征量计算表达式;以及评价值计算装置,配置成计算在所述特征量提取表达式列表中包括的每个特征量提取表达式的所述评价值。 And a feature amount and the actual data supplied from the teaching data corresponding to the existing feature amount, as taught by the data used to estimate the actual supply data corresponding to the target machine learning feature quantity to generate the feature quantity calculation target expression; and an evaluation value calculating means configured to calculate each feature amount extraction expression list extraction expressions included in the evaluation value feature quantity.
2.根据权利要求1所述的信息处理设备,其中所述目标特征量计算表达式生成装置同等地且有选择地利用所计算的与所述实际数据对应的所述多个特征量中的一些特征量和与作为教导数据供应的所述实际数据对应的多个现有特征量中的一些特征量,以通过用于估计与作为教导数据供应的所述实际数据对应的所述目标特征量的机器学习来生成所述目标特征量计算表达式。 The information processing apparatus according to claim 1, wherein the target feature quantity calculation expression to generate some of the plurality of feature amounts corresponding to the real data and apparatus are equally selective use of the calculated feature amount and a feature amount as a number of said plurality of existing feature amount corresponding to the actual data supplied from the teaching data in order for the target feature by estimating the teaching data supplied as the actual data corresponding to an amount of machine learning to generate the target feature amount calculation expression.
3.根据权利要求1所述的信息处理设备,其中所述评价值计算装置基于所计算的与所述实际数据对应的所述多个特征量的所述目标特征量计算表达式的贡献率来计算所述特征量提取表达式列表中包括的每个特征量提取表达式的所述评价值。 The information processing apparatus according to claim 1, wherein the contribution ratio calculating means calculates the value of an expression based on the calculated target feature amount of the feature amounts of the plurality of actual calculated data corresponding Review each feature amount extraction expression list of the feature amount extraction expressions included in the evaluation value.
4. 一种信息处理设备,用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,所述信息处理设备包括:特征量提取表达式列表生成装置,配置成将在前一代的包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于所述特征量提取表达式的评价值的遗传算法更新前一代的所述特征量提取表达式列表来生成所述特征量提取表达式列表;特征量计算装置,配置成将作为教导数据供应的实际数据输入到在所述特征量提取表达式列表中包括的每个特征量提取表达式以计算与所述实际数据对应的多个特征量、并且也测量各个特征量提取表达式的平均计算时间;目标特征量计算表达式生成装置,配置成利用所计算的与所述实际数据对应的所述多个特征量以通过用于估计 4. An information processing apparatus for generating a target feature amount corresponding to the target input data and the feature amount calculation expression, said information processing apparatus comprising: feature amount extraction expression list generating means, configured to in a plurality of feature amounts include a plurality of feature amounts of a plurality of operators composed extraction expression previous generation feature quantity extraction expression list included in the expression for gene extraction, the extracted feature based on the assessment of the expression amount by using updating the value of the genetic algorithm for feature amount extraction expression lists of the previous generation to generate the feature amount extraction expression list; feature quantity calculating means configured to input the actual data to the teaching data supplied from the feature amount extraction each feature amount extraction expression list including expression to calculate the actual data corresponding to a plurality of feature amounts, and also measuring the respective feature amount extraction expression average computation time; target feature amount calculation expression generating means, configured to utilize the calculated and the actual data corresponding to the plurality of feature amounts for estimating by 作为教导数据供应的所述实际数据对应的目标特征量的机器学习来生成所述目标特征量计算表达式;以及评价值计算装置,配置成计算在所述特征量提取表达式列表中包括的每个特征量提取表达式的所述评价值、并且也基于各个特征量提取表达式的所述平均计算时间来校正所述计算的评价值。 Corresponding target feature amount as the teaching data supplied from the actual data to generate the target machine learning feature quantity calculation expressions; and an evaluation value calculating means configured to calculate the feature amount extraction expression lists each including a feature amount extraction expression evaluation value, and also based on the respective feature amount extraction expressions of the average computation time correcting said calculated evaluation value.
5.根据权利要求4所述的信息处理设备,其中所述目标特征量计算表达式生成装置有选择地利用所计算的与所述实际数据对应的所述多个特征量中的一些特征量以通过用于估计与作为教导数据供应的所述实际数据对应的所述目标特征量的机器学习来生成所述目标特征量计算表达式。 The information processing apparatus according to claim 4, wherein the target feature quantity calculation expression generating means has a number of said plurality of feature quantity to the feature quantity corresponding to the actual data using the calculated selectively in order to by estimating a teaching data as the actual data supplied from the feature quantity corresponding to the target machine learning to generate the target feature amount calculation expression.
6.根据权利要求5所述的信息处理设备,其中所述目标特征量计算表达式生成装置基于所述对应特征量提取表达式的所述平均计算时间有选择地利用所计算的与所述实际数据对应的所述多个特征量中的一些特征量,以通过用于估计与作为教导数据供应的所述实际数据对应的所述目标特征量的机器学习来生成所述目标特征量计算表达式。 The information processing apparatus according to claim 5, wherein the target feature quantity calculating the average computation time expression generating means for extracting a feature amount based on the corresponding expression are selected and the actual use of the calculated Some machine learning feature data corresponding to the plurality of feature amounts, through for estimating the feature quantity as the target of the teaching data corresponding to the actual data supplied from the target to generate the feature quantity calculation expression .
7. 一种信息处理设备,用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,所述信息处理设备包括:特征量提取表达式列表生成装置,配置成将在前一代的包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表中包括的多个特征量提取表达式作为基因,通过使用基于所述特征量提取表达式的评价值的遗传算法更新前一代的所述特征量提取表达式列表来生成所述特征量提取表达式列表;特征量计算装置,配置成将作为教导数据供应的实际数据输入到在所述特征量提取表达式列表中包括的每个特征量提取表达式以计算与所述实际数据对应的多个特征量;目标特征量计算表达式生成装置,配置成利用所计算的与所述实际数据对应的所述多个特征量以通过用于估计与作为教导数据供应的所述实际数据对应的目标特征 An information processing apparatus for generating a target feature amount corresponding to the target input data and the feature amount calculation expression, said information processing apparatus comprising: feature amount extraction expression list generating means, configured to in a plurality of feature amounts include a plurality of feature amounts by a plurality of operators composed extraction expression previous generation feature quantity extraction expression list including gene expression as extraction, extraction expression-based assessment by using the feature quantity updating the value of the genetic algorithm for feature amount extraction expression lists of the previous generation to generate the feature amount extraction expression list; feature quantity calculating means configured to input the actual data to the teaching data supplied from the feature amount extraction expression of each feature amount extraction expression list including a plurality of computing the feature amount corresponding to the actual data; target feature amount calculation expression generating means configured to utilize the calculated data corresponding to the actual through said plurality of feature amounts for estimating the real data corresponding to a target characteristic data supplied from the teachings 的机器学习来生成所述目标特征量计算表达式;评价值计算装置,配置成计算在所述特征量提取表达式列表中包括的每个特征量提取表达式的所述评价值;以及优化装置,配置成优化在最后一代的所述特征量提取表达式列表中包括的多个特征量提取表达式中的每个特征量提取表达式。 Machine learning to generate the feature quantity calculation target expression; evaluation value calculating means configured to calculate each feature amount extraction expression list extraction expressions included in the evaluation value of the feature amount; and optimization means, configured to optimize a plurality of feature amount extraction expression list including each extracted feature amount extraction expression in the expression of the last generation of the feature amount.
8.根据权利要求7所述的信息处理设备,其中所述优化装置包括:特征量提取表达式优化装置,配置成:通过从最后一代的所述特征量提取表达式列表中包括的各个特征量提取表达式中检测表示了预先登记的冗余运算符的组合的优化模式并且通过删除运算符或者替换为算术负荷较小的运算符,来优化在最后一代的所述特征量提取表达式列表中包括的每个特征量提取表达式。 Feature amount extraction expression optimizing means is configured to:: by extraction from various features of the last generation of the feature amount included in the expression list according to the amount of the information processing apparatus according to claim 7, wherein said optimizing means comprises extraction expression detected pattern represents the combination of optimized redundancy operators registered in advance and by deleting or replacing the arithmetic operator smaller load operator to optimize the extraction in the last generation of the feature amount list of expressions each feature amount extraction expressions included.
9.根据权利要求7所述的信息处理设备,其中所述优化装置包括:特征量提取表达式优化装置,配置成:变形在最后一代的所述特征量提取表达式列表中包括的每个特征量提取表达式以生成多个优化候选表达式;向所述多个生成的优化候选表达式之中的以下优化候选表达式赋予优良评价:即获得的该优化候选表达式的输出具有与作为变形源的所述特征量提取表达式的输出的高相关度、并且该优化候选表达式的计算时间较短;将所述多个生成的优化候选表达式作为基因,利用基于所述优化候选表达式的评价的遗传算法以更新所述多个生成的优化候选表达式;以及将具有最优良评价的所述优化候选表达式最终确定为在最后一代的所述特征量提取表达式列表中包括的各个特征量提取表达式的优化结果。 9. The information processing apparatus according to claim 7, wherein said optimizing means comprising: a feature amount extraction expression optimizing means is configured to: extract each modified expression features included in the list of the final generation feature quantity amount extraction expression to generate a plurality of candidate expressions optimization; optimization candidate among the plurality expression produces the following expression gives good candidates for optimization Rating: optimization of the expression of the candidate has the output that is obtained as a modification the feature amount extraction expression source outputs a high correlation, and the optimization calculation time shorter candidate expression; the plurality of generated gene expression as candidates for optimization, the optimization based on candidate using expression to update said plurality of candidate expressions to generate optimized evaluation of the genetic algorithm; having the best evaluation and optimization of the respective candidate is determined as the final expression extraction expression lists in the final generation of the feature quantity comprises the feature amount extraction expression optimization results.
10.根据权利要求7所述的信息处理设备,其中所述优化装置包括:重构装置,配置成利用所述优化的特征量提取表达式以重构与最后一代的所述特征量提取表达式列表对应生成的所述目标特征量计算表达式。 10. The information processing apparatus according to claim 7, wherein said optimizing means comprising: a reconstruction means configured to, with the feature amount extraction expression optimized to reconstruct the last generation of the feature amount extraction expressions the target generation feature quantity calculation expression corresponding to the list.
11. 一种信息处理方法,用于获取输入数据和与所述输入数据对应的现有特征量作为输入、并且生成用于输出与所述输入数据对应的目标特征量的目标特征量计算表达式,所述信息处理方法包括以下步骤:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在所述特征量提取表达式列表中包括的每个特征量提取表达式,以计算与所述实际数据对应的多个特征量;同等地利用所计算的与所述实际数据对应的所述多个特征量和与作为教导数据供应的所述实际数据对应的现有特征量,以通过用于估计与作为教导数据供应的所述实际数据对应的目标特征量的机器学习来生成所述目标特征量计算表达式;计算在所述特征量提取表达式列表中包括的每个特征量提取表达式的评价值;以及将在前 11. An information processing method for acquiring the input data and the input data corresponding to the conventional feature amount as an input and generates a feature amount of the target output and the input data corresponding to the target feature amount calculation expression for said information processing method comprising the steps of: randomly generating a plurality of feature amounts of a plurality of operators composed feature amount extraction expression list extraction expression; the actual data inputted as the teaching data supplied to the feature quantity each feature amount extraction expression list including extraction expression to calculate the actual data corresponding to a plurality of feature amounts; equally calculated using the plurality of the feature amount and the actual data and the corresponding the target feature, as the teaching data supplied from the data corresponding to the actual existing feature quantity, through a teaching data for estimating the actual supply amount of the data corresponding to the target machine learning to generate the feature quantity calculation expression; calculating the feature quantity of each feature amount extraction expression list including extraction expression evaluation value; and front 一代的所述特征量提取表达式列表中包括的多个特征量提取表达式作为基因, 通过使用基于所述特征量提取表达式的所述评价值的遗传算法更新前一代的所述特征量提取表达式列表来生成下一代的所述特征量提取表达式列表。 A plurality of feature amounts of the feature amount extraction expression list generation includes extraction expression gene, the feature amount extraction expression evaluation value of the genetic algorithm updates the previous generation by using the extracted feature amount based on the expression of the list is generated feature quantity extraction expression lists of the next generation.
12. 一种信息处理方法,用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,所述信息处理方法包括以下步骤:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在所述特征量提取表达式列表中包括的每个特征量提取表达式以计算与所述实际数据对应的多个特征量、并且也测量各个特征量提取表达式的平均计算时间;利用所计算的与所述实际数据对应的所述多个特征量以通过用于估计与作为教导数据供应的所述实际数据对应的目标特征量的机器学习来生成所述目标特征量计算表达式;计算在所述特征量提取表达式列表中包括的每个特征量提取表达式的评价值、并且也基于各个特征量提取表达式的所述平均计算时间来校正所述计算的评价值;以及将在前一代 12. An information processing method for generating a target feature amount corresponding to the target input data and the feature amount calculation expression, said information processing method comprising the steps of: randomly generating a plurality of operators composed of a plurality a feature amount extraction expression list of feature amount extraction expression; the actual input data as teaching data supplied to the feature amount extraction expression lists each including extraction expression to calculate the actual data with the feature amount a corresponding plurality of feature amounts, and also measuring the respective feature amount extraction expression of the average computation time; using said plurality of feature amounts calculated and the actual data corresponding to the estimated by serving as a teaching data supplied said actual data corresponding to the target feature amount generating the machine learning target feature amount calculation expression; calculating for each feature amount extraction expression list including extraction expression evaluation value of the feature amount, and also based on respective the average time is calculated feature amount extraction expression to calculate the corrected evaluation value; and first generation 的所述特征量提取表达式列表中包括的多个特征量提取表达式作为基因, 通过使用基于所述特征量提取表达式的所述评价值的遗传算法更新前一代的所述特征量提取表达式列表来生成下一代的所述特征量提取表达式列表。 The characteristic amount extraction expression list of the plurality of feature amounts include extraction expression gene, the feature amount extraction expression evaluation value of the genetic algorithm updates the previous generation based on the feature quantity extracted by using the expression generating a list of the next generation of the feature amount extraction expression list.
13. 一种信息处理方法,用于生成用于输出与输入数据对应的目标特征量的目标特征量计算表达式,所述信息处理方法包括以下步骤:随机生成包括由多个运算符组成的多个特征量提取表达式的特征量提取表达式列表;将作为教导数据供应的实际数据输入到在所述特征量提取表达式列表中包括的每个特征量提取表达式以计算与所述实际数据对应的多个特征量;利用所计算的与所述实际数据对应的所述多个特征量以通过用于估计与作为教导数据供应的所述实际数据对应的目标特征量的机器学习来生成所述目标特征量计算表达式;计算在所述特征量提取表达式列表中包括的每个特征量提取表达式的评价值;将在前一代的所述特征量提取表达式列表中包括的多个特征量提取表达式作为基因, 通过使用基于所述特征量提取表达式的所述评价值的遗传算法更 13. An information processing method for generating a target feature amount corresponding to the target input data and the feature amount calculation expression, said information processing method comprising the steps of: randomly generating a plurality of operators composed of a plurality a feature amount extraction expression list of feature amount extraction expression; the actual input data as teaching data supplied to the feature amount extraction expression lists each including extraction expression to calculate the actual data with the feature amount corresponding to the plurality of feature amounts; machine learning using the plurality of feature amounts calculated and the actual data corresponding to a via for estimating the actual data supplied from the teaching data corresponding to the target feature amount to generate said target feature amount calculation expression; calculating feature amount extraction expression lists each including extraction expression evaluation value of the feature amount; and extracting the feature amount of the first generation expression list comprises a plurality of feature amount extraction for gene expression, genetic algorithms extraction expression evaluation value based on the feature quantity by using a more 前一代的所述特征量提取表达式列表来生成下一代的所述特征量提取表达式列表;以及优化在最后一代的所述特征量提取表达式列表中包括的多个特征量提取表达式中的每个特征量提取表达式。 The feature amount of the feature amount extraction expression list generation before generating a next extraction expression list; and extracting a plurality of feature amounts optimizing expression included in the list of the extracted feature quantity of the last generation expression each feature amount extraction expression.
CN 200810170287 2007-10-22 2008-10-20 Information processing device, information processing method CN101419610B (en)

Priority Applications (6)

Application Number Priority Date Filing Date Title
JP2007273416A JP4392621B2 (en) 2007-10-22 2007-10-22 Information processing apparatus, information processing method, and program
JP2007-273417 2007-10-22
JP2007-273418 2007-10-22
JP2007273417A JP4433323B2 (en) 2007-10-22 2007-10-22 Information processing apparatus, information processing method, and program
JP2007273418A JP4392622B2 (en) 2007-10-22 2007-10-22 Information processing apparatus, information processing method, and program
JP2007-273416 2007-10-22

Publications (2)

Publication Number Publication Date
CN101419610A CN101419610A (en) 2009-04-29
CN101419610B true CN101419610B (en) 2011-11-16

Family

ID=40630402

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 200810170287 CN101419610B (en) 2007-10-22 2008-10-20 Information processing device, information processing method

Country Status (2)

Country Link
JP (1) JP4433323B2 (en)
CN (1) CN101419610B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5581864B2 (en) * 2010-07-14 2014-09-03 ソニー株式会社 Information processing apparatus, information processing method, and program
JP5881048B2 (en) * 2012-09-18 2016-03-09 株式会社日立製作所 Information processing system and information processing method
CN106708609A (en) * 2015-11-16 2017-05-24 阿里巴巴集团控股有限公司 Characteristics generation method and system

Also Published As

Publication number Publication date
JP2009104274A (en) 2009-05-14
JP4433323B2 (en) 2010-03-17
CN101419610A (en) 2009-04-29

Similar Documents

Publication Publication Date Title
Yang et al. A regression approach to music emotion recognition
Ryynänen et al. Automatic transcription of melody, bass line, and chords in polyphonic music
Mauch et al. Simultaneous estimation of chords and musical context from audio
Cemgil et al. On tempo tracking: Tempogram representation and Kalman filtering
Turnbull et al. A Game-Based Approach for Collecting Semantic Annotations of Music.
Gillet et al. Transcription and separation of drum signals from polyphonic music
US7754958B2 (en) Sound analysis apparatus and program
Chandna et al. Monoaural audio source separation using deep convolutional neural networks
Schorfheide DSGE model-based estimation of the New Keynesian Phillips curve
Yang et al. Ranking-based emotion recognition for music organization and retrieval
Vincent et al. The signal separation evaluation campaign (2007–2010): Achievements and remaining challenges
CN101116134B (en) Information processing device, method and program
Durrieu et al. Source/filter model for unsupervised main melody extraction from polyphonic audio signals
Benetos et al. Automatic music transcription: challenges and future directions
JPWO2010053160A1 (en) Content search apparatus and content search program
Stamatatos et al. Automatic identification of music performers with learning ensembles
US8706274B2 (en) Information processing apparatus, information processing method, and program
US7908135B2 (en) Music-piece classification based on sustain regions
CN103635960A (en) Statistical enhancement of speech output from statistical text-to-speech synthesis system
Schmidt et al. Prediction of Time-varying Musical Mood Distributions from Audio.
Raffel Learning-based methods for comparing sequences, with applications to audio-to-midi alignment and matching
Nam et al. A Classification-Based Polyphonic Piano Transcription Approach Using Learned Feature Representations.
Schmidt et al. Learning emotion-based acoustic features with deep belief networks
Benetos et al. A shift-invariant latent variable model for automatic music transcription
US8112418B2 (en) Generating audio annotations for search and retrieval

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C14 Grant of patent or utility model
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20111116

Termination date: 20181020