JP4121060B2 - Class identification device and class identification method - Google Patents

Description

【００５０】
このように、個々の観測データの分類への所属状態から、分類結果評価値を算出することにより、種類判別子付き観測データが少なく、種類判別子の有無にかかわらず、全観測データを分類対象とするような場合にも評価が可能となり、分類対象データの分布状況を反映した分類を得ることができる。
【００５１】
分類再実行指示生成器４では、中間結果評価器５から出力される中間結果評価値を第一の入力とし、分類パラメータファイル７２から出力される分類パラメータを第二の入力として、分類再実行の必要性を判断し、分類再実行パラメータを出力する。判断の仕方としては、あらかじめ設定されていた閾値と中間結果評価値とを比較し、閾値以下の場合に再実行とする。閾値は、分類パラメータの一つとしてユーザが設定するか、あるいは、中間結果の評価方法に応じて、あらかじめシステムが決めておく。
【００５２】
次に、分類再実行パラメータの例として、分類再実行指示生成器４を構成する分類パラメータ変更器４１において、分類パラメータを変更して分類パラメータ変更値を求める場合を示す。分類パラメータとしては、乱数のシード値や分類手法に特化したパラメータ等があるが、ここでは分類数を変更する場合の例を示す。
【００５３】
ユーザが変更値を指定して分類数を変更する方法もあるが、以下では、システムが自動的に変更を行う例を示す。例えば、分類数が極端に多い場合や、少ない場合に、観測データ数に見合った分類数になるように変更する。なお、観測データ数は分類パラメータとして与えられているものとする。
【００５４】
具体的には、観測データ数の１／Ｎｃｄを分類数変更値とする。ここで、Ｎｃｄは１分類あたり属する平均観測データ数であり、１以上で観測データ数以下の正の整数とする。このＮｃｄは、システムが乱数を使って決定するか、あるいは分類パラメータとしてユーザが指定してもよい。たとえば、観測データ数が１００でＮｃｄ＝２０の場合、１００／２０＝５が分類数となる。
【００５５】
このように、初期設定として与えられた分類数などの分類パラメータを変更して、分類し直すことにより、より良い分類結果が期待できる。
【００５６】
分類再実行指示生成器４から出力された分類再実行パラメータは、観測データ分類器１の入力となり、再び、観測データ分類器１において分類及び分類結果の出力を行い、続いて中間結果評価器５において評価及び中間結果評価値の出力を行う。以下、分類再実行指示生成器４において分類再実行不要と判断されるまで、観測データ分類器１、中間結果評価器５、及び分類再実行指示生成器４において上記の処理を繰り返す。分類再実行指示生成器４において、分類再実行不要と判断されたところで、分類終了指示が分類再実行パラメータとして出力される。観測データ分類器１では、分類再実行指示生成器４から出力された分類終了指示が入力されると、観測データ分類結果を種類学習器２に出力する。
【００５７】
分類再実行指示生成器４は、中間結果評価値が事前に設定されていたある閾値以上になるか、あるいは事前に設定されていたある回数を超えたところで、再実行不要と判断する。また、観測データ分類器１は、上記繰り返し処理で得られた観測データ分類結果のうち、最良の分類結果評価値を記録した観測データ分類結果を出力する。
【００５８】
次に、種類学習器２において、学習パラメータファイル７１から出力される学習パラメータを第一の入力とし、種類判別子付き観測データファイル９から出力される上記種類判別子付き観測データを第二の入力とし、観測データ分類器１から出力される観測データ分類結果を第三の入力として、上記観測データ分類結果で示される各分類内で、上記種類判別子付き観測データに基づき種類学習を行い、種類学習の結果である種類学習結果を出力する。
【００５９】
種類学習を行う例を示す。本実施の形態１では、学習手法として境界線を決定する手法である線形判別法を適用する。この線形判別法は、全教師データからの距離と方向に基づき境界となる直線を求める方法であり、距離を最大にし、かつそれぞれの種類ごとに、教師データからみた直線の方向が同じになるような直線を決定する。
【００６０】
図２の中央の図に線形判別法による種類学習例を示す。『△』は種類１の教師データ、『黒□』は種類２の教師データであるとすると、種類学習により、図２の中央の図のように境界線が決定される。
【００６１】
種類判別器３では、種類学習器２から出力される種類学習結果を第一の入力とし、種類判別対象観測データファイル１０から出力される種類判別対象観測データを第二の入力として、上記種類学習結果に基づき上記種類判別対象観測データの種類判別を行い、結果を種類判別結果として出力する。この出力は、本発明の類識別装置６Ａ全体の出力となり、種類判別結果ファイル１１に入力される。
【００６２】
図２の右図に種類判別を行う例を示す。種類学習結果として求められた境界線に基づき、種類を判別する。図２の右図で『△』が多く属する境界線の上側の領域を種類１の境界領域、逆に『黒□』が多く属する境界線の下側の領域を種類２の境界領域とし、いずれの領域に属するかによって種類判別対象観測データの種類を判別する。図２の右図において、『黒▽』と『◇』は種類が未定の観測データであり、『黒▽』と『◇』がそれぞれ種類１、種類２と判別される。
【００６３】
以上のように、本発明の実施の形態１で示した類識別装置によれば、個々の中間結果を評価でき、個々の中間結果の状況に応じて、分類段階からやり直すことができる。
【００６４】
また、分類パラメータ変更器４１において、初期設定として与えられた分類数などの分類パラメータを変更して、分類し直すことにより、より良い分類結果が期待できる。
【００６５】
また、分類結果評価器５１において、分類段階で評価を行って分類し直すことにより、より良い分類結果に基づいて学習を行うことができる。
【００６６】
また、分類結果評価器５１において、種類判別子付き観測データの種類ごとの分類への所属状態から、分類結果評価値を算出することにより、多種種類を対象とした場合に、分類段階である程度、種類分けを行うような評価基準を導入することができ、効率の良い学習が期待できる。
【００６７】
さらに、分類結果評価器５１において、個々の観測データの分類への所属状態から、分類結果評価値を算出することにより、種類判別子付き観測データが少なく、種類判別子の有無にかかわらず、全観測データを分類対象とするような場合にも評価が可能となり、分類対象データの分布状況を反映した分類を得ることができる。
【００６８】
実施の形態２．
この発明の実施の形態２に係る類識別装置及び類識別方法について図面を参照しながら説明する。図３は、この発明の実施の形態２に係る類識別装置の構成を示す図である。
【００６９】
図３において、１は観測データ分類器、２は種類学習器、３は種類判別器、４は分類再実行指示生成器、５は中間結果評価器、６Ｂは類識別装置、７１は学習パラメータファイル、７２は分類パラメータファイル、９は種類判別子付き観測データファイル、１０は種類判別対象観測データファイル、１１は種類判別結果ファイルである。また、分類再実行指示生成器４は、分類手法変更器４２を有する。
【００７０】
つぎに、この実施の形態２に係る類識別装置の動作について図面を参照しながら説明する。
【００７１】
上記実施の形態１では、分類再実行指示生成器４を構成する分類パラメータ変更器４１により分類パラメータを変更して分類再実行を行う例を示したが、本実施の形態２では、分類再実行指示生成器４を構成する分類手法変更器４２によって分類手法を変更して分類再実行を行う例を示す。
【００７２】
第一の分類手法変更例として、上記実施の形態１で示した最短距離法から、Ｋ−ｍｅａｎｓ法に変更する場合の例を以下に示す。
【００７３】
このＫ−ｍｅａｎｓ法とは、一般的な分類問題に多く用いられている手法である。個々の観測データをｋ個の分類に適当に割り当て、重心の初期値を決定する。次に、個々の観測データは最も近い重心を持つ分類に属するものとし、各分類において、属する観測データを使って重心を求め直し、さらに更新された重心に基づき、個々の観測データが属する分類を変更するといった処理を何度か繰り返し、最終的な分類を決定する。この場合の分類パラメータとしては、重心の初期値を求める際に、各観測データが属する分類を決定するための乱数のシード値などがある。
【００７４】
Ｋ−ｍｅａｎｓ法と前述の最短距離法との違いは、最短距離法が分類を階層的に行うのに対し、Ｋ−ｍｅａｎｓ法では非階層的に行われる点である。最短距離法のように階層的に分類を行う場合、分類対象となる観測データ数が多い場合に、分類処理に多くの時間を費やしてしまうが、Ｋ−ｍｅａｎｓ法による非階層的分類では、最初から分類を初期設定し、分類の評価と改良を繰り返して最終的な分類を決定するため、繰り返し回数の上限指定による実行時間の制御が可能である。したがって、分類対象となる観測データ数が多い場合であっても、効率の良い分類が期待できる。
【００７５】
このように、新たな分類手法として、Ｋ−ｍｅａｎｓ法を適用することにより、分類対象となる観測データ数が多い場合等に、効率よく分類を行うことができる。
【００７６】
第二の分類手法変更例として、分類手法をＫ−ｍｅａｎｓ法からＦｕｚｚｙ ｃ−ｍｅａｎｓ法に変更する場合の例を以下に示す。
【００７７】
このＦｕｚｚｙ ｃ−ｍｅａｎｓ法は、Ｋ−ｍｅａｎｓ法の変形手法の１つである。Ｋ−ｍｅａｎｓ法では、各観測データがいずれの分類に属するかを一意に決定するが、Ｆｕｚｚｙ ｃ−ｍｅａｎｓ法では、それぞれの分類への所属確率で表す。
【００７８】
例えば、分類ｄ１〜分類ｄ４の４つの分類に対するデータ１の所属確率Ｐ１（ｄｉ）とすると、Ｐ１（ｄ１）＝０．１、Ｐ１（ｄ２）＝０．５、Ｐ１（ｄ３）＝０．１５、Ｐ１（ｄ４）＝０．２５のように表される。なお、所属確率Ｐ１（ｄｉ）は、メンバーシップ関数と呼ばれる関数で定義される。
【００７９】
ここで、ある観測データが、２つの分類（分類１、分類２）のどちらに所属しても良いような場合を考える。例えば、２つの分類に対し、Ｐ（１）＝０．４５、Ｐ（２）＝０．５５程度の所属確率が予想される場合、Ｋ−ｍｅａｎｓ法では、必ず一方に分類を決定するため、わずかに確率の高い分類２に分類される可能性が高い。これは実質、Ｐ（１）＝０、Ｐ（２）＝１とみなしたのと同じであり、このような処理を繰り返すうちに、多くの誤差が生じて、最終的な分類精度が悪くなる危険性が高い。
【００８０】
そこで、このような問題に対しては、Ｆｕｚｚｙ ｃ−ｍｅａｎｓ法での所属確率による分類が有効であり、精度の高い分類が期待できる。また、Ｆｕｚｚｙｃ−ｍｅａｎｓ法のメンバーシップ関数の定義にガウス分布を導入することにより、期待値最大化法と呼ばれるガウス分布でデータの分布状況を定義する手法への変更も可能となる。
【００８１】
このように、新たな分類手法としてＦｕｚｚｙ ｃ−ｍｅａｎｓ法を適用することにより、分類への所属が曖昧な観測データが存在する場合に、精度の高い分類が可能になる。
【００８２】
以上のように、本発明の実施の形態２で示した類識別装置によれば、分類手法変更器４２において、問題の性質に応じて、分類手法を変更することにより、より良い分類結果が期待できる。
【００８３】
また、同じく分類手法変更器４２において、新たな分類手法としてＫ−ｍｅａｎｓ法を適用することにより、分類対象となる観測データ数が多い場合等に、効率よく分類を行うことができる。
【００８４】
さらに、同じく分類手法変更器４２において、新たな分類手法としてＦｕｚｚｙ ｃ−ｍｅａｎｓ法を適用することにより、分類への所属が曖昧な観測データが存在する場合に、精度の高い分類が可能になる。
【００８５】
実施の形態３．
この発明の実施の形態３に係る類識別装置及び類識別方法について図面を参照しながら説明する。図４は、この発明の実施の形態３に係る類識別装置の構成を示す図である。
【００８６】
図４において、１は観測データ分類器、２は種類学習器、３は種類判別器、４は分類再実行指示生成器、５は中間結果評価器、６Ｃは類識別装置、７１は学習パラメータファイル、７２は分類パラメータファイル、９は種類判別子付き観測データファイル、１０は種類判別対象観測データファイル、１１は種類判別結果ファイルである。また、中間結果評価器５は、学習結果評価器５２を有する。
【００８７】
つぎに、この実施の形態３に係る類識別装置の動作について図面を参照しながら説明する。
【００８８】
図５及び図６は、この発明の実施の形態３に係る類識別装置の動作を示す図である。
【００８９】
上記実施の形態１では、中間結果評価器５を構成する分類結果評価器５１において分類結果を評価する例を示したが、本実施の形態３では、同じく中間結果評価器５を構成する学習結果評価器５２において学習結果を評価する例を示す。
【００９０】
学習結果評価方法の第一の例としては、学習対象外の種類判別子付き観測データの帰属状態から、正解率を算出し、学習結果評価値とする。
【００９１】
具体的には、学習対象外とした種類判別子付き観測データに対して、種類判別を行い、その誤認率を算出して学習結果評価値とする。例えば、図５の右図において、『黒▽』、『◇』を学習対象外とした種類判別子付き観測データであるとし、それぞれ種類１、種類２であるとすると、１２個中１０個が正しく判別され、２個が誤った種類に判別されている。これより、学習評価結果値は、Ｅ１１＝１０／１２＝０．８３となる。なお、０≦Ｅｌ１≦１．０であるとする。
【００９２】
以上のように、学習に影響していない学習対象外の種類判別子付き観測データの、上記学習結果で示される境界領域への所属状態から、正解率を算出することにより、学習対象データの偏りに依存することなく、評価を行うことができる。
【００９３】
学習結果評価方法の第二の例としては、学習対象としたデータを含む全ての種類判別子付き観測データの所属状態から、誤認の度合いを算出し、学習結果評価値とする。
【００９４】
具体的には、学習対象としたデータを含む全ての種類判別子付き観測データに対して、種類判別を行い、誤って判別されたデータの境界線からの距離に基づき学習評価値を算出する。
【００９５】
つまり、まず、全学習対象データの境界線からの距離の平均値ｍを求める。次に、誤認されたデータｉの境界線からの距離ｄｉをｍで割った値ｄｉ／ｍを誤認距離ｄｎｉとし、誤認距離の総和Σｄｎｉを誤認データ数Ｎｍで割った値であるΣｄｎｉ／Ｎｍを学習結果評価値とする。
【００９６】
例えば、図６において、『△』、『黒□』を学習に適用した種類判別子付き観測データ、『黒▽』、『◇』を学習の対象外の種類判別子付き観測データであるとすると、１９個中３個が誤った種類の境界領域に位置する。
【００９７】
今、誤認したデータと境界線との距離がそれぞれｄ１＝２、ｄ２＝２、ｄ３＝５で、全学習対象データの境界線からの距離の平均値ｍ＝１０、誤認データ数Ｎｍ＝３とすると、誤認距離はそれぞれｄｎ１＝０．２、ｄｎ２＝０．２、ｄｎ３＝０．５となる。したがって、学習評価結果値はＥｌ１＝Σｄｎｉ／Ｎｍ＝（０．２＋０．２＋０．５）／３＝０．３となる。但し、０≦Ｅｌ１とする。なお、この場合、学習評価値は０が最良であり、値が大きいほど評価値は悪くなる。
【００９８】
学習対象としたデータを含む全ての上記種類判別子付き観測データのうち、誤認であると判定されたデータに対し、上記学習結果で示される境界線からの距離を用いて、誤認の度合いを示す誤認距離を算出し、上記誤認距離の総和に基づき上記学習結果評価値を算出することにより、種類判別子付き観測データが少なく、全ての種類判別子付き観測データを学習対象とする場合にも評価可能となる。
【００９９】
学習結果評価値は、分類再実行指示生成器４の入力となり、上記実施の形態１と同様な処理で分類が変更され、変更された新たな観測データ分類結果は再び種類学習器２の入力となり、以下、評価と変更を繰り返し、学習結果評価値がある閾値以上になるか、あるいはある回数以上になった場合に、繰り返しで得られた最良の学習結果評価値を記録した学習結果を、種類判別器３の入力となる種類学習結果として出力する。
【０１００】
以上のように、本発明の実施の形態３で示した類識別装置６Ｃによれば、学習結果評価器５２において、種類学習結果の評価を行い、その種類学習結果を観測データ分類器１にフィードバックすることにより、種類学習結果を反映した分類を行うことができ、これは類識別処理全体の質の向上につながり、結果として質の高い種類判別を期待できる。
【０１０１】
また、同じく学習結果評価器５２において、学習に影響していない学習対象外の種類判別子付き観測データの上記学習結果で示される境界領域への所属状態から、正解率を算出することにより、学習対象データの偏りに依存することなく、評価を行うことができる。
【０１０２】
さらに、同じく学習結果評価器５２において、学習対象としたデータを含む全ての上記種類判別子付き観測データのうち、誤認であると判定されたデータに対し、上記学習結果で示される境界線からの距離を用いて、誤認の度合いを示す誤認距離を算出し、上記誤認距離の総和に基づき上記学習結果評価値を算出することにより、種類判別子付き観測データが少なく、全ての種類判別子付き観測データを学習対象とする場合にも評価可能となる。
【０１０３】
なお、本発明は航空機の類識別を行う場合の例を示したが、それ以外にも、レーダから得られた大量の電波画像データを利用した類識別問題への適用も可能である。
【０１０９】
【発明の効果】
この発明の請求項１に係る類識別装置は、以上説明したとおり、種類を表す種類判別子付きの観測データである種類判別子付き観測データを最短距離法により分類し、収束条件として、分類の数をあらわす分類数になったところで、分類を決定する観測データ分類器と、（分類ｊに所属する種類ｉのデータ数）／（種類ｉの全データ数）から前記種類判別子付き観測データの種類ｉの分類ｊにおける所属確率Ｐ _ｉｊ を算出し、種類ｉにおける最大所属確率をＰ _ｉｍａｘ 、種類数をＮｋとすると、ΣＰ _ｉｍａｘ ／Ｎｋから分類結果評価値Ｅｋを算出する分類結果評価器と、前記分類結果評価値Ｅｋと所定の閾値とを比較し、前記分類結果評価値Ｅｋが前記所定の閾値以下の場合には分類再実行が必要であると判断した場合は、Ｎｃｄを１分類あたり属する平均観測データ数とすると、分類数を、観測データ数の１／Ｎｃｄと変更して出力し、分類再実行が不要と判断した場合には、分類終了指示を出力する分類パラメータ変更器と、前記観測データ分類器により分類された種類判別子付き全観測データからの距離と方向に基づき境界となる直線を求める線形判別法により種類学習を行う種類学習器と、分類毎の境界線の上側、下側のいずれの領域に属するかによって種類判別対象観測データの種類を判別する種類判別器とを備え、前記分類パラメータ変更器において分類再実行が不要と判断され前記分類終了指示が出力されるまで、前記観測データ分類器において分類及び分類結果の出力を行い、続いて前記分類結果評価器において評価及び分類結果評価値の出力を行う処理を繰り返し、前記観測データ分類器では、前記分類終了指示が入力されると、繰り返し処理で得られた観測データ分類結果のうち、最良の分類結果評価値を記録した観測データ分類結果を前記種類学習器に出力するので、多種種類を対象とした場合に、分類段階である程度、種類分けを行うような評価基準を導入することができ、効率の良い学習が期待できるという効果を奏する。
【０１１０】
この発明の請求項２に係る類識別装置は、以上説明したとおり、種類を表す種類判別子付きの観測データである種類判別子付き観測データを最短距離法により分類し、収束条件として、分類の数をあらわす分類数になったところで、分類を決定する観測データ分類器と、種類判別子の有無にかかわらない観測データｍの各分類への所属確率が最大の値である、観測データｍにおける最大所属確率をＰ_ｍｍａｘ、データ数をＮｄとすると、ΣＰ_ｍｍａｘ／Ｎｄから分類結果評価値Ｅｄを算出する分類結果評価器と、前記分類結果評価値Ｅｄと所定の閾値とを比較し、前記分類結果評価値Ｅｄが前記所定の閾値以下の場合には分類再実行が必要であると判断した場合は、Ｎｃｄを１分類あたり属する平均観測データ数とすると、分類数を、観測データ数の１／Ｎｃｄと変更して出力し、分類再実行が不要と判断した場合には、分類終了指示を出力する分類パラメータ変更器と、前記観測データ分類器により分類された種類判別子付き全観測データからの距離と方向に基づき境界となる直線を求める線形判別法により種類学習を行う種類学習器と、分類毎の境界線の上側、下側のいずれの領域に属するかによって種類判別対象観測データの種類を判別する種類判別器とを備え、前記分類パラメータ変更器において分類再実行が不要と判断され前記分類終了指示が出力されるまで、前記観測データ分類器において分類及び分類結果の出力を行い、続いて前記分類結果評価器において評価及び分類結果評価値の出力を行う処理を繰り返し、前記観測データ分類器では、前記分類終了指示が入力されると、繰り返し処理で得られた観測データ分類結果のうち、最良の分類結果評価値を記録した観測データ分類結果を前記種類学習器に出力するので、種類判別子付き観測データが少なく、種類判別子の有無にかかわらず、全観測データを分類対象とするような場合にも評価が可能となり、分類対象データの分布状況を反映した分類を得ることができるという効果を奏する。
【０１１６】
この発明の請求項３に係る類識別方法は、以上説明したとおり、観測データ分類器により、種類を表す種類判別子付きの観測データである種類判別子付き観測データを最短距離法により分類し、収束条件として、分類の数をあらわす分類数になったところで、分類を決定する観測データ分類ステップと、分類結果評価器により、（分類ｊに所属する種類ｉのデータ数）／（種類ｉの全データ数）から前記種類判別子付き観測データの種類ｉの分類ｊにおける所属確率Ｐ_ｉｊを算出し、種類ｉにおける最大所属確率をＰ_ｉｍａｘ、種類数をＮｋとすると、ΣＰ_ｉｍａｘ／Ｎｋから分類結果評価値Ｅｋを算出する分類結果評価ステップと、分類パラメータ変更器により、前記分類結果評価値Ｅｋと所定の閾値とを比較し、前記分類結果評価値Ｅｋが前記所定の閾値以下の場合には分類再実行が必要であると判断した場合は、Ｎｃｄを１分類あたり属する平均観測データ数とすると、分類数を、観測データ数の１／Ｎｃｄと変更して出力し、分類再実行が不要と判断した場合には、分類終了指示を出力する分類パラメータ変更ステップと、種類学習器により、前記観測データ分類ステップにより分類された種類判別子付き全観測データからの距離と方向に基づき境界となる直線を求める線形判別法により種類学習を行う種類学習ステップと、種類判別器により、分類毎の境界線の上側、下側のいずれの領域に属するかによって種類判別対象観測データの種類を判別する種類判別ステップとを含み、前記分類パラメータ変更ステップにおいて分類再実行が不要と判断され前記分類終了指示が出力されるまで、前記観測データ分類ステップにおいて分類及び分類結果の出力を行い、続いて前記分類結果評価ステップにおいて評価及び分類結果評価値の出力を行う処理を繰り返し、前記観測データ分類ステップでは、前記分類終了指示が入力されると、繰り返し処理で得られた観測データ分類結果のうち、最良の分類結果評価値を記録した観測データ分類結果を前記種類学習ステップに出力するので、分類段階で評価を行って分類し直すことにより、より良い分類結果に基づいて学習を行うことができるという効果を奏する。
【０１１７】
この発明の請求項４に係る類識別方法は、以上説明したとおり、観測データ分類器により、種類を表す種類判別子付きの観測データである種類判別子付き観測データを最短距離法により分類し、収束条件として、分類の数をあらわす分類数になったところで、分類を決定する観測データ分類ステップと、分類結果評価器により、種類判別子の有無にかかわらない観測データｍの各分類への所属確率が最大の値である、観測データｍにおける最大所属確率をＰ_ｍｍａｘ、データ数をＮｄとすると、ΣＰ_ｍｍａｘ／Ｎｄから分類結果評価値Ｅｄを算出する分類結果評価ステップと、分類パラメータ変更器により、前記分類結果評価値Ｅｄと所定の閾値とを比較し、前記分類結果評価値Ｅｄが前記所定の閾値以下の場合には分類再実行が必要であると判断した場合は、Ｎｃｄを１分類あたり属する平均観測データ数とすると、分類数を、観測データ数の１／Ｎｃｄと変更して出力し、分類再実行が不要と判断した場合には、分類終了指示を出力する分類パラメータ変更ステップと、種類学習器により、前記観測データ分類ステップにより分類された種類判別子付き全観測データからの距離と方向に基づき境界となる直線を求める線形判別法により種類学習を行う種類学習ステップと、種類判別器により、分類毎の境界線の上側、下側のいずれの領域に属するかによって種類判別対象観測データの種類を判別する種類判別ステップとを含み、前記分類パラメータ変更ステップにおいて分類再実行が不要と判断され前記分類終了指示が出力されるまで、前記観測データ分類ステップにおいて分類及び分類結果の出力を行い、続いて前記分類結果評価ステップにおいて評価及び分類結果評価値の出力を行う処理を繰り返し、前記観測データ分類ステップでは、前記分類終了指示が入力されると、繰り返し処理で得られた観測データ分類結果のうち、最良の分類結果評価値を記録した観測データ分類結果を前記種類学習ステップに出力するので、分類段階で評価を行って分類し直すことにより、より良い分類結果に基づいて学習を行うことができるという効果を奏する。
【図面の簡単な説明】
【図１】 この発明の実施の形態１に係る類識別装置の構成を示す図である。
【図２】 この発明の実施の形態１に係る類識別装置の動作を示す図である。
【図３】 この発明の実施の形態２に係る類識別装置の構成を示す図である。
【図４】 この発明の実施の形態３に係る類識別装置の構成を示す図である。
【図５】 この発明の実施の形態３に係る類識別装置の動作を示す図である。
【図６】 この発明の実施の形態３に係る類識別装置の動作を示す図である。
【図７】 従来の類識別装置の構成を示す図である。
【図８】 従来の類識別装置の動作を示す図である。
【符号の説明】
１ 観測データ分類器、２ 種類学習器、３ 種類判別器、４ 分類再実行指示生成器、４１ 分類パラメータ変更器、４２ 分類手法変更器、５ 中間結果評価器、５１ 分類結果評価器、５２ 学習結果評価器、６Ａ、６Ｂ、６Ｃ 類識別装置、７１ 学習パラメータファイル、７２ 分類パラメータファイル、９種類判別子付き観測データファイル、１０ 種類判別対象観測データファイル、１１ 種類判別結果ファイル。[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a class identification problem for determining the type of a class identification object, and more particularly to a class identification device and a class identification method for obtaining the type of the class identification object as a solution.
[0002]
[Prior art]
As an example of the class identification problem, consider an aircraft class identification problem.
[0003]
A conventional class identification device will be described with reference to the drawings. FIG. 7 is shown in, for example, “A trial of maintenance training support using object recognition by Support Vector Machine” written by Yoshito Nakajima, Operations Research, May 2001, pages 237 (15) to 242 (20). It is a figure which shows the structure of the conventional class identification apparatus which performs class identification of the made aircraft.
[0004]
In FIG. 7, 2 is a type learner, 3 is a type discriminator, 6 is a conventional class discriminator, 9 is an observation data file with a type discriminator, 71 is a learning parameter file, 10 is a type discrimination target observation data file, 11 Is a type discrimination result file.
[0005]
Next, the operation of the conventional class identification device will be described with reference to the drawings.
[0006]
FIG. 8 is a diagram illustrating the operation of a conventional class identification device.
[0007]
First, in the type learning device 2 of the class identification device 6, the learning parameter output from the learning parameter file 71 is the first input, and the observation data with type discriminator output from the observation data file with type discriminator 9 is the first. As a second input, type learning is performed based on the observation data with the type discriminator, and a type learning result that is a result of the type learning is output.
[0008]
As a conventional learning method for determining the boundary line, there is a linear discrimination method. This linear discriminant method is a method of finding a straight line that is a boundary based on the distance and direction from all the teacher data. The distance is maximized, and the direction of the straight line seen from the teacher data is the same for each type. Determine a straight line.
[0009]
The left figure of FIG. 8 shows an example of boundary determination by the linear discriminant method. Assuming that “Δ” is type 1 teacher data and “black □” is type 2 teacher data, the boundary line is determined as shown in the left diagram of FIG. In the drawing, the black square is described as “black □” because it cannot be used in electronic applications. The same applies to “Black”.
[0010]
In the type discriminator 3 of the class discriminating device 6, the type learning result output from the type learning device 2 is used as a first input, and the type discrimination target observation data output from the type discrimination target observation data file 10 is input as a second input. As a result, the type of the type discrimination target observation data is discriminated based on the type learning result, and the result is output as the type discrimination result. This output is the output of the entire conventional class identification device 6 and is input to the type discrimination result file 11.
[0011]
An example of performing type discrimination is shown in the right diagram of FIG. The type is determined based on the boundary line obtained as the type learning result. In the right diagram of FIG. 8, the type discrimination target observation data belonging to the region including “Δ” is discriminated as type 1, and the type discrimination target observation data belonging to the region including “black □” is discriminated as type 2. Now, the type determination is performed for the data “black ▽”, “◇”, and “*” whose types are undecided. “Black” and “◇” are data belonging to Type 1 and Type 2, respectively, and “*” is an unknown type of data that is neither Type 1 nor Type 2. In the right diagram of FIG. 8, many “black ▽” and “◇” are correctly determined as type 1 and type 2, respectively, but some are erroneously determined. In addition, “*” is determined to be type 2 even though it is an unknown type.
[0012]
From the above, it can be seen that the conventional class identification device 6 has a high possibility that there is data that cannot be correctly identified in the vicinity of the boundary line. When such an erroneous discrimination result is obtained, generally, there are many cases where the whole process is re-executed from the beginning using new data with a type discriminator. However, there may be a case where the data cannot be changed, for example, when the data with type discriminator is not abundant.
[0013]
[Problems to be solved by the invention]
The conventional class identification device as described above has a problem that there is a high possibility that there is data that cannot be correctly identified in the vicinity of the boundary line.
[0014]
The present invention has been made to solve the above-described problems. When the accuracy of the type discrimination result is poor, the type discrimination accuracy is not changed without changing the data with the type discriminator, which is data for type learning. An object of the present invention is to obtain a class identification device and a class identification method capable of increasing
[0015]
[Means for Solving the Problems]
  A class identification device according to claim 1 of the present invention provides:Classify the observation data with type discriminator, which is the observation data with type discriminator that represents the type, by the shortest distance method, and determine the classification when the convergence number is the number of classifications representing the number of classificationsAn observation data classifier;From the (number of data of type i belonging to class j) / (total number of data of type i), the membership probability P in class j of type i of the observation data with type discriminator _ij And the maximum affiliation probability in type i is P _imax If the number of types is Nk, ΣP _imax The classification result evaluator that calculates the classification result evaluation value Ek from / Nk, the classification result evaluation value Ek and a predetermined threshold value are compared, and if the classification result evaluation value Ek is less than or equal to the predetermined threshold value, the classification If it is determined that execution is necessary, assuming that Ncd is the average number of observation data belonging to one class, the number of classification is changed to 1 / Ncd of the number of observation data, and it is determined that classification re-execution is unnecessary. In this case, a classification parameter changer that outputs a classification end instruction and a type of linear discriminant that obtains a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classifier. Do learningA kind learner,Determine the type of observation data to be classified according to whether it belongs to the area above or below the boundary line for each classification.Type discriminator andUntil the classification parameter changer determines that classification re-execution is unnecessary and outputs the classification end instruction, and outputs the classification and classification result in the observed data classifier, and then in the classification result evaluator When the observation data classification unit receives the classification end instruction, the best classification result evaluation value is obtained from the observation data classification results obtained by the repetition process. Output the observation data classification result that recordsIs.
[0016]
  The class identification device according to claim 2 of the present invention classifies observation data with a type discriminator, which is observation data with a type discriminator representing a type, by the shortest distance method, and the classification number representing the number of classifications as a convergence condition The observation data classifier that determines the classification,Observations that have the highest affiliation probability for each classification of observation data m regardless of the presence or absence of a type discriminatorThe maximum membership probability in data m is P_mmaxIf the number of data is Nd, ΣP_mmaxThe classification result evaluator that calculates the classification result evaluation value Ed from / Nd, the classification result evaluation value Ed and a predetermined threshold value are compared, and if the classification result evaluation value Ed is equal to or less than the predetermined threshold value, the classification If it is determined that execution is necessary, assuming that Ncd is the average number of observation data belonging to one class, the number of classification is changed to 1 / Ncd of the number of observation data, and it is determined that classification re-execution is unnecessary. In this case, a classification parameter changer that outputs a classification end instruction and a type of linear discriminant that obtains a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classifier. A classification learner that performs learning, and a type discriminator that discriminates the type of type discrimination target observation data depending on whether it belongs to an area above or below the boundary line for each classification, and the classification parameter Until the reclassifier determines that classification re-execution is unnecessary and outputs the classification end instruction, the observation data classifier outputs the classification and the classification result, and then the classification result evaluator evaluates the classification result and the classification result evaluation value. In the observation data classifier, when the classification end instruction is input, the observation data classification in which the best classification result evaluation value is recorded among the observation data classification results obtained by the repetition processing is input to the observation data classifier The result is output to the kind learning device.
[0017]
  The class identification method according to claim 3 of the present invention includes:Observation data classifierObservation data classification step that classifies observation data with a type discriminator, which is an observation data with a type discriminator that represents the type, by the shortest distance method, and determines the classification when the number of classifications represents the number of classifications as the convergence condition When,Classification result evaluatorFrom the (number of data of type i belonging to class j) / (total number of data of type i), the membership probability P in class j of type i of the observation data with type discriminator_ijAnd the maximum affiliation probability in type i is P_imaxIf the number of types is Nk, ΣP_imaxA classification result evaluation step of calculating a classification result evaluation value Ek from / Nk;Classification parameter changerWhen the classification result evaluation value Ek is compared with a predetermined threshold, and it is determined that the classification re-execution is necessary when the classification result evaluation value Ek is equal to or less than the predetermined threshold, Ncd belongs to one classification If the average number of observation data is assumed, the classification number is changed to 1 / Ncd of the observation data number and is output. If it is determined that classification re-execution is unnecessary, a classification parameter changing step for outputting a classification end instruction;By type learning device,A type learning step for performing type learning by a linear discriminant method for obtaining a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classification step;By type discriminator,A type determining step for determining the type of the type determining target observation data depending on whether the region belongs to the upper side or the lower side of the boundary line for each classification, and the classification parameter changing step determines that classification re-execution is unnecessary. Until the classification end instruction is output, the process of outputting the classification and the classification result in the observation data classification step, and then repeating the process of outputting the evaluation and the classification result evaluation value in the classification result evaluation step is repeated. In the step, when the classification end instruction is input, the observation data classification result in which the best classification result evaluation value is recorded is output to the type learning step among the observation data classification results obtained by the iterative process. .
[0018]
  The class identification method according to claim 4 of the present invention includes:Observation data classifierObservation data classification step that classifies observation data with a type discriminator, which is an observation data with a type discriminator that represents the type, by the shortest distance method, and determines the classification when the number of classifications represents the number of classifications as the convergence condition When,Observation with classification result evaluator that has the maximum value of belonging to each classification of observation data m regardless of the presence or absence of type discriminatorThe maximum membership probability in data m is P_mmaxIf the number of data is Nd, ΣP_mmaxA classification result evaluation step of calculating a classification result evaluation value Ed from / Nd;Classification parameter changerThe classification result evaluation value Ed is compared with a predetermined threshold, and if the classification result evaluation value Ed is equal to or less than the predetermined threshold and it is determined that classification re-execution is necessary, Ncd belongs to one classification. If the average number of observation data is assumed, the classification number is changed to 1 / Ncd of the observation data number and is output. If it is determined that classification re-execution is unnecessary, a classification parameter changing step for outputting a classification end instruction;By type learning device,A type learning step for performing type learning by a linear discriminant method for obtaining a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classification step;By type discriminator,A type determining step for determining the type of the type determining target observation data depending on whether the region belongs to the upper side or the lower side of the boundary line for each classification, and the classification parameter changing step determines that classification re-execution is unnecessary. Until the classification end instruction is output, the process of outputting the classification and the classification result in the observation data classification step, and then repeating the process of outputting the evaluation and the classification result evaluation value in the classification result evaluation step is repeated. In the step, when the classification end instruction is input, the observation data classification result in which the best classification result evaluation value is recorded is output to the type learning step among the observation data classification results obtained by the iterative process. .
[0031]
DETAILED DESCRIPTION OF THE INVENTION
Embodiment 1 FIG.
A class identification device and a class identification method according to Embodiment 1 of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing a configuration of a class identification device according to Embodiment 1 of the present invention. In addition, in each figure, the same code | symbol shows the same or equivalent part.
[0032]
In FIG. 1, 1 is an observation data classifier, 2 is a type learner, 3 is a type discriminator, 4 is a classification re-execution instruction generator, 5 is an intermediate result evaluator, 6A is a class identification device, and 71 is a learning parameter file. 72 is a classification parameter file, 9 is an observation data file with a type discriminator, 10 is a type discrimination target observation data file, and 11 is a type discrimination result file. Further, the classification re-execution instruction generator 4 has a classification parameter changer 41, and the intermediate result evaluator 5 has a classification result evaluator 51.
[0033]
Next, the operation of the class identification device according to the first embodiment will be described with reference to the drawings.
[0034]
FIG. 2 is a diagram showing the operation of the class identification device according to Embodiment 1 of the present invention.
[0035]
First, in the observation data classifier 1, a classification parameter that is a parameter for classification output from the classification parameter file 72 as an input of the entire class identification device 6A is used as a first input, and also as an input of the entire class identification device 6A. The observation data with a type discriminator, which is the observation data with a type discriminator representing the type, output from the observation data file 9 with a type discriminator is used as the second input, and is a classification result of the observation data with the type discriminator. Output the observation data classification result. The classification parameters include, for example, the number of classifications representing the number of classifications, random seed values for determining initial values of parameters necessary for classification, parameters specialized for classification methods, and the like.
[0036]
An example of classification of observation data with type discriminator is shown. In the first embodiment, a nearest neighbor method, which is a method often used for general classification problems, is applied as a classification method. In the shortest distance method, as an initial setting, each observation data is regarded as one independent class, and the similarity between the classes is represented by the maximum value of the similarity between arbitrary observation data belonging to each class. Next, the classifications having the highest similarity are combined, and the similarity is calculated again. The following is a method of determining classification when the combination of classification and calculation of similarity are repeated and the convergence condition is satisfied. Convergence conditions include that the similarity between all classifications is equal to or less than a preset threshold, and that the number of classifications is also set in advance. The similarity between observation data is calculated from the degree of matching of attribute values. The classification parameter in this case includes a similarity threshold. The left side of FIG. 2 shows the result of classification for 15 observation data of “△” and “Black □”.
[0037]
In the intermediate result evaluator 5, the intermediate result output from the type learning device 2 or the observation data classifier 1 is used as the first input, and the type output from the type discrimination target observation data file 10 as the input of the entire class identification device 6A. The discrimination target observation data is used as a second input, an intermediate result output from any of the devices constituting the class identification device 6A is evaluated, and the result is output as an intermediate result evaluation value. Here, an example in which the classification result is evaluated in the classification result evaluator 51 constituting the intermediate result evaluator 5 is shown.
[0038]
Specifically, paying attention to the affiliation state of the observation data with type discriminator to the classification, it is determined whether the classification to which the data belongs can be limited to some extent for each type. For each type, the affiliation probability in the classification to which the most data belongs is obtained, and the classification result evaluation value is calculated from the sum of the affiliation probabilities. Here, when the classification to which the observation data with the type discriminator belongs is uniquely determined as in the classification by the shortest distance method, the number of belonging data in each classification for each type is counted.
[0039]
Now, it is assumed that the observation data with a type discriminator is composed of three types of 22 data, and the breakdown is 9 for type 1, 6 for type 2, and 7 for type 3. Classification results,
Type 1 has 2 in Category 1, 7 in Category 2,
Type 2 is 5 in Category 1, 1 in Category 2,
It is assumed that type 3 is divided into two for category 1 and five for category 2.
[0040]
Affiliation probability P in classification i of type i_ijThe
P_ij= (Number of data of type i belonging to category j) / (Total number of data of type i)
And calculating
P₁₁= 2/9 = 0.22,
P₁₂= 7/9 = 0.78,
P₂₁= 5/6 = 0.83,
P₂₂= 1/6 = 0.17,
P₃₁= 2/7 = 0.28,
P₃₂= 5/7 = 0.72.
[0041]
Next, let P be the maximum membership probability for type i._imaxThen,
P_1max= 0.78, P_2max= 0.83, P_3max= 0.72, and the sum of these sums divided by the number of types Nk = 3 is used as the classification result evaluation value Ek. In the case of the above example, it is as follows. Ek takes a value of 1 / Nc ≦ Ek ≦ 1.0 (Nc is the number of classifications).
[0042]
Ek = ΣP_imax/Nk=(0.78+0.83+0.72)/3=0.78
[0043]
In the above example, the case where the observation data with each type discriminator is uniquely assigned to the classification is shown. However, depending on the classification method, it may not be uniquely determined but may be given probabilistically. Even in such a case, the classification result evaluation value is calculated as in the case of the shortest distance method, assuming that each data belongs to the classification with the highest probability.
[0044]
In this way, by calculating the classification result evaluation value from the affiliation state to the classification for each type of data with a type discriminator, the classification is performed to some extent at the classification stage when targeting various types. Evaluation criteria can be introduced and efficient learning can be expected.
[0045]
Next, a calculation example of the second classification result evaluation value is shown. In this example, a case where the affiliation to the classification of the observation data including the type discrimination target observation data is not determined uniquely but is given probabilistically is targeted. The classification result evaluation value is calculated from the average value of the maximum affiliation probabilities to the classification of the observation data regardless of the presence or absence of the type discriminator.
[0046]
Here, if the affiliation probabilities of data 1 in category 1, category 2, and category 3 are 0.1, 0.5, and 0.4, the maximum value of 0.5 is assigned to data 1 as the maximum affiliation. Probability. The maximum membership probability in data m is P_mmaxThen, the result obtained by dividing these sums by the number of data Nd is the classification result evaluation value Ed.
[0047]
Now, the observation data of the number of data Nd = 7 is targeted, and their maximum affiliation probability P_mmaxIs P_1max= 0.7, P_2max= 0.8, P_3max= 0.5, P_4max= 0.9, P_5max= 0.4, P_6max= 0.7, P_7max= 0.6. The number of classifications is 3.
[0048]
In the case of the above example, it is as follows. Note that Ed takes a value of 1 / Nc ≦ Ed ≦ 1.0 (Nc is the number of classifications).
[0049]

[0050]
In this way, by calculating the classification result evaluation value from the affiliation status of individual observation data, there are few observation data with type discriminators, and all observation data are classified regardless of whether or not there is a type discriminator. In such a case, the evaluation can be performed, and the classification reflecting the distribution status of the classification target data can be obtained.
[0051]
The classification re-execution instruction generator 4 uses the intermediate result evaluation value output from the intermediate result evaluator 5 as a first input and the classification parameter output from the classification parameter file 72 as a second input. Judge the necessity and output the classification re-execution parameter. As a determination method, a preset threshold value is compared with the intermediate result evaluation value, and if it is equal to or less than the threshold value, re-execution is performed. The threshold is set by the user as one of the classification parameters or determined in advance by the system according to the evaluation method of the intermediate results.
[0052]
Next, as an example of the classification re-execution parameter, a case where the classification parameter change unit 41 constituting the classification re-execution instruction generator 4 changes the classification parameter to obtain the classification parameter change value will be described. As classification parameters, there are random seed values, parameters specialized for classification methods, and the like. Here, an example of changing the number of classifications is shown.
[0053]
Although there is a method in which the user specifies the change value and changes the number of classifications, an example in which the system automatically changes will be described below. For example, when the number of classifications is extremely large or small, the number of classifications is changed to match the number of observation data. It is assumed that the number of observation data is given as a classification parameter.
[0054]
Specifically, 1 / Ncd of the number of observation data is set as the classification number change value. Here, Ncd is the average number of observation data belonging to one class, and is a positive integer greater than or equal to 1 and less than or equal to the number of observation data. This Ncd may be determined by the system using a random number, or may be specified by the user as a classification parameter. For example, when the number of observation data is 100 and Ncd = 20, 100/20 = 5 is the classification number.
[0055]
In this way, a better classification result can be expected by changing classification parameters such as the number of classifications given as the initial setting and reclassifying.
[0056]
The classification re-execution parameter output from the classification re-execution instruction generator 4 becomes the input of the observation data classifier 1, and the observation data classifier 1 again outputs the classification and the classification result, and then the intermediate result evaluator 5 The evaluation and the intermediate result evaluation value are output at. Thereafter, the above processing is repeated in the observation data classifier 1, the intermediate result evaluator 5, and the classification re-execution instruction generator 4 until the classification re-execution instruction generator 4 determines that classification re-execution is unnecessary. When the classification re-execution instruction generator 4 determines that classification re-execution is unnecessary, a classification end instruction is output as a classification re-execution parameter. The observation data classifier 1 outputs the observation data classification result to the type learning device 2 when the classification end instruction output from the classification re-execution instruction generator 4 is input.
[0057]
The classification re-execution instruction generator 4 determines that re-execution is not necessary when the intermediate result evaluation value is equal to or greater than a predetermined threshold value or exceeds a predetermined number of times. The observation data classifier 1 outputs an observation data classification result in which the best classification result evaluation value is recorded among the observation data classification results obtained by the above iterative processing.
[0058]
Next, in the type learning device 2, the learning parameter output from the learning parameter file 71 is a first input, and the observation data with a type discriminator output from the observation data file with a type discriminator 9 is a second input. Using the observation data classification result output from the observation data classifier 1 as the third input, the type learning is performed based on the observation data with the type discriminator within each classification indicated by the observation data classification result. The kind learning result which is a learning result is output.
[0059]
An example of performing type learning is shown. In the first embodiment, a linear discriminant method that is a method for determining a boundary line is applied as a learning method. This linear discriminant method is a method of finding a straight line that is a boundary based on the distance and direction from all the teacher data. The distance is maximized, and the direction of the straight line seen from the teacher data is the same for each type. Determine a straight line.
[0060]
An example of type learning by the linear discriminant method is shown in the center diagram of FIG. Assuming that “Δ” is type 1 teacher data and “black □” is type 2 teacher data, the boundary line is determined by type learning as shown in the center diagram of FIG.
[0061]
The type discriminator 3 uses the type learning result output from the type learner 2 as a first input and the type discriminating target observation data output from the type discriminating target observation data file 10 as a second input. Based on the result, type discrimination of the type discrimination target observation data is performed, and the result is output as a type discrimination result. This output becomes the output of the entire class identification device 6A of the present invention and is input to the type discrimination result file 11.
[0062]
An example of performing type discrimination is shown in the right diagram of FIG. The type is determined based on the boundary line obtained as the type learning result. In the right figure of FIG. 2, the area above the boundary line to which many “△” belongs is the type 1 boundary area, and conversely, the area below the boundary line to which many “black squares” belong is the type 2 boundary area. The type of the type discrimination target observation data is discriminated depending on whether it belongs to the region. In the right diagram of FIG. 2, “black ▽” and “◇” are observation data whose types are undetermined, and “black ▽” and “◇” are determined as type 1 and type 2, respectively.
[0063]
As described above, according to the class identification device shown in the first embodiment of the present invention, each intermediate result can be evaluated, and it can be redone from the classification stage according to the situation of each intermediate result.
[0064]
In the classification parameter changer 41, a better classification result can be expected by changing classification parameters such as the number of classifications given as an initial setting and reclassifying the classification parameters.
[0065]
Further, the classification result evaluator 51 can perform learning based on a better classification result by performing evaluation and reclassifying at the classification stage.
[0066]
Further, in the classification result evaluator 51, by calculating the classification result evaluation value from the belonging state to the classification for each type of the observation data with a type discriminator, in the case of targeting various types, to some extent at the classification stage. Evaluation criteria for classifying can be introduced, and efficient learning can be expected.
[0067]
Furthermore, the classification result evaluator 51 calculates the classification result evaluation value from the belonging state of each observation data to the classification, so that there are few observation data with a type discriminator, and all the data regardless of the presence or absence of the type discriminator. Evaluation is possible even when the observation data is a classification target, and a classification reflecting the distribution status of the classification target data can be obtained.
[0068]
Embodiment 2. FIG.
A class identification device and a class identification method according to Embodiment 2 of the present invention will be described with reference to the drawings. FIG. 3 is a diagram showing a configuration of a class identification device according to Embodiment 2 of the present invention.
[0069]
In FIG. 3, 1 is an observation data classifier, 2 is a type learner, 3 is a type discriminator, 4 is a classification re-execution instruction generator, 5 is an intermediate result evaluator, 6B is a class identification device, and 71 is a learning parameter file. 72 is a classification parameter file, 9 is an observation data file with a type discriminator, 10 is a type discrimination target observation data file, and 11 is a type discrimination result file. Further, the classification re-execution instruction generator 4 has a classification technique changer 42.
[0070]
Next, the operation of the class identification device according to the second embodiment will be described with reference to the drawings.
[0071]
In the first embodiment, the example in which the classification parameter is changed by the classification parameter changer 41 constituting the classification re-execution instruction generator 4 and the classification is re-executed is shown. In the second embodiment, the classification re-execution is performed. An example in which the classification technique is changed by the classification technique changer 42 constituting the instruction generator 4 and classification is re-executed will be described.
[0072]
As a first classification method change example, an example in which the shortest distance method shown in the first embodiment is changed to the K-means method will be described below.
[0073]
The K-means method is a technique that is often used for general classification problems. Individual observation data is appropriately assigned to k classifications, and an initial value of the center of gravity is determined. Next, it is assumed that each observation data belongs to the category with the nearest centroid, and in each classification, the centroid is re-determined using the observation data to which it belongs, and the classification to which the individual observation data belongs is based on the updated centroid. The process of changing is repeated several times to determine the final classification. The classification parameter in this case includes a random seed value for determining the classification to which each observation data belongs when determining the initial value of the center of gravity.
[0074]
The difference between the K-means method and the shortest distance method described above is that the shortest distance method performs classification hierarchically, whereas the K-means method performs non-hierarchical. When performing hierarchical classification as in the shortest distance method, if the number of observation data to be classified is large, a lot of time is spent on the classification process. However, in the non-hierarchical classification by the K-means method, Since the classification is initially set and the final classification is determined by repeating the evaluation and improvement of the classification, it is possible to control the execution time by specifying the upper limit of the number of repetitions. Therefore, efficient classification can be expected even when the number of observation data to be classified is large.
[0075]
As described above, by applying the K-means method as a new classification method, classification can be performed efficiently when the number of observation data to be classified is large.
[0076]
As a second example of changing the classification method, an example in which the classification method is changed from the K-means method to the Fuzzy c-means method is shown below.
[0077]
The Fuzzy c-means method is one of the modified methods of the K-means method. In the K-means method, which classification each observation data belongs to is uniquely determined, but in the Fuzzy c-means method, it is represented by the belonging probability to each classification.
[0078]
For example, assuming that the affiliation probability P1 (di) of data 1 for the four classifications d1 to d4 is P1 (d1) = 0.1, P1 (d2) = 0.5, P1 (d3) = 0.15. , P1 (d4) = 0.25. The membership probability P1 (di) is defined by a function called a membership function.
[0079]
Here, a case is considered in which a certain observation data may belong to either of two classifications (classification 1 and classification 2). For example, when the affiliation probabilities of P (1) = 0.45 and P (2) = 0.55 are predicted for two classifications, the K-means method always determines the classification to one of There is a high possibility of being classified into category 2 with a slightly high probability. This is substantially the same as assuming that P (1) = 0 and P (2) = 1, and many errors occur and the final classification accuracy deteriorates as this process is repeated. High risk.
[0080]
Therefore, for such a problem, classification based on the affiliation probability in the Fuzzy c-means method is effective, and high-precision classification can be expected. In addition, by introducing a Gaussian distribution into the definition of the membership function of the Fuzzyc-means method, it is possible to change to a method for defining a data distribution state with a Gaussian distribution called an expected value maximization method.
[0081]
Thus, by applying the Fuzzy c-means method as a new classification method, it is possible to classify with high accuracy when there is observation data whose affiliation to the classification is ambiguous.
[0082]
As described above, according to the class identification device shown in the second exemplary embodiment of the present invention, a better classification result can be expected by changing the classification method according to the nature of the problem in the classification method changer 42. it can.
[0083]
Similarly, by applying the K-means method as a new classification method in the classification method changer 42, the classification can be performed efficiently when the number of observation data to be classified is large.
[0084]
Further, similarly, by applying the Fuzzy c-means method as a new classification method in the classification method changer 42, it is possible to classify with high accuracy when there is observation data whose affiliation to the classification is ambiguous.
[0085]
Embodiment 3 FIG.
A class identification device and class identification method according to Embodiment 3 of the present invention will be described with reference to the drawings. FIG. 4 is a diagram showing a configuration of a class identification device according to Embodiment 3 of the present invention.
[0086]
In FIG. 4, 1 is an observation data classifier, 2 is a type learner, 3 is a type discriminator, 4 is a classification re-execution instruction generator, 5 is an intermediate result evaluator, 6C is a class identification device, and 71 is a learning parameter file. 72 is a classification parameter file, 9 is an observation data file with a type discriminator, 10 is a type discrimination target observation data file, and 11 is a type discrimination result file. Further, the intermediate result evaluator 5 has a learning result evaluator 52.
[0087]
Next, the operation of the class identification device according to the third embodiment will be described with reference to the drawings.
[0088]
5 and 6 are diagrams showing the operation of the class identification device according to Embodiment 3 of the present invention.
[0089]
In the first embodiment, the example in which the classification result is evaluated by the classification result evaluator 51 that constitutes the intermediate result evaluator 5 has been shown. However, in the third embodiment, the learning result that also constitutes the intermediate result evaluator 5 is shown. An example in which the evaluator 52 evaluates the learning result is shown.
[0090]
As a first example of the learning result evaluation method, a correct answer rate is calculated from the belonging state of observation data with a type discriminator that is not a learning target, and is used as a learning result evaluation value.
[0091]
Specifically, type discrimination is performed on observation data with a type discriminator that is not subject to learning, and the misperception rate is calculated to obtain a learning result evaluation value. For example, in the right diagram of FIG. 5, assuming that “black ▽” and “◇” are observation data with a type discriminator and are type 1 and type 2, respectively, 10 out of 12 It is determined correctly and two are determined to be of the wrong type. Thus, the learning evaluation result value is E11 = 10/12 = 0.83. It is assumed that 0 ≦ El1 ≦ 1.0.
[0092]
As described above, the bias of the learning target data is calculated by calculating the accuracy rate from the affiliation state of the observation data with a type discriminator that is not affected by learning and does not affect learning to the boundary region indicated by the learning result. The evaluation can be performed without depending on.
[0093]
As a second example of the learning result evaluation method, the misidentification degree is calculated from the affiliation state of all the observation data with type discriminators including the data to be learned, and set as a learning result evaluation value.
[0094]
Specifically, type discrimination is performed on all observation data with a type discriminator including data to be learned, and a learning evaluation value is calculated based on the distance from the boundary line of the data discriminated incorrectly.
[0095]
That is, first, an average value m of distances from the boundary lines of all learning target data is obtained. Next, a value di / m obtained by dividing the distance di from the boundary line of the misidentified data i by m is defined as the misidentified distance dni, and the sum Σdni of the misidentified distance is divided by the number Nm of misidentified data Σdni / Nm. Let it be a learning result evaluation value.
[0096]
For example, in FIG. 6, suppose that “Δ” and “black □” are observation data with a type discriminator applied to learning, and “black ▽” and “◇” are observation data with a type discriminator not subject to learning. , 3 out of 19 are located in the wrong type of boundary region.
[0097]
Now, the distance between the misidentified data and the boundary line is d1 = 2, d2 = 2, and d3 = 5, respectively, the average value m = 10 from the boundary line of all learning target data, and the number of misidentified data Nm = 3. Then, the misidentification distances are dn1 = 0.2, dn2 = 0.2, and dn3 = 0.5, respectively. Therefore, the learning evaluation result value is El1 = Σdni / Nm = (0.2 + 0.2 + 0.5) /3=0.3. However, 0 ≦ El1. In this case, 0 is the best learning evaluation value, and the larger the value, the worse the evaluation value.
[0098]
Shows the degree of misidentification using the distance from the boundary indicated by the learning result for the data determined to be misidentified among all the observation data with type discriminators including the data to be learned By calculating the misidentification distance and calculating the learning result evaluation value based on the sum of the misidentification distances, evaluation is possible even when there are few observation data with type discriminators and all observation data with type discriminators are targeted for learning. It becomes possible.
[0099]
The learning result evaluation value is input to the classification re-execution instruction generator 4, the classification is changed by the same processing as in the first embodiment, and the new observation data classification result thus changed is input to the type learning device 2 again. After that, the evaluation and change are repeated, and when the learning result evaluation value exceeds a certain threshold or exceeds a certain number of times, the learning result that records the best learning result evaluation value obtained by repetition is classified into The result is output as a type learning result to be input to the discriminator 3.
[0100]
As described above, according to the class identification device 6C shown in the third embodiment of the present invention, the learning result evaluator 52 evaluates the type learning result and feeds back the type learning result to the observation data classifier 1. By doing so, it is possible to perform classification that reflects the type learning result, which leads to improvement in the quality of the entire class identification process, and as a result, high-quality type discrimination can be expected.
[0101]
Similarly, the learning result evaluator 52 calculates the correct answer rate from the affiliation state to the boundary region indicated by the learning result of the observation data with the type discriminator that is not subject to learning and does not affect the learning. The evaluation can be performed without depending on the bias of the target data.
[0102]
Furthermore, in the learning result evaluator 52, among all the observation data with the type discriminator including the data to be learned, the data determined to be misidentified is from the boundary indicated by the learning result. By using the distance, the misidentification distance indicating the degree of misperception is calculated, and the learning result evaluation value is calculated based on the sum of the misidentification distances. Evaluation is possible even when data is a learning target.
[0103]
Although the present invention shows an example of classifying an aircraft, the present invention can be applied to a class identification problem using a large amount of radio wave image data obtained from a radar.
[0109]
【The invention's effect】
  Of this inventionClaim 1As described above, the class identification device according toAn observation data classifier that classifies observation data with a type discriminator, which is an observation data with a type discriminator representing a type, by the shortest distance method, and determines the classification when the number of classifications represents the number of classifications as a convergence condition. And (the number of data of type i belonging to class j) / (the total number of data of type i), the belonging probability P in class j of type i of the observation data with type discriminator _ij And the maximum affiliation probability in type i is P _imax If the number of types is Nk, ΣP _imax The classification result evaluator that calculates the classification result evaluation value Ek from / Nk, the classification result evaluation value Ek and a predetermined threshold value are compared, and if the classification result evaluation value Ek is less than or equal to the predetermined threshold value, the classification If it is determined that execution is necessary, assuming that Ncd is the average number of observation data belonging to one class, the number of classification is changed to 1 / Ncd of the number of observation data, and it is determined that classification re-execution is unnecessary. In this case, a classification parameter changer that outputs a classification end instruction and a type of linear discriminant that obtains a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classifier. A classification learner that performs learning, and a type discriminator that discriminates the type of type discrimination target observation data depending on whether it belongs to an area above or below the boundary line for each classification, and the classification parameter Until the reclassifier determines that classification re-execution is unnecessary and outputs the classification end instruction, the observation data classifier outputs the classification and the classification result, and then the classification result evaluator evaluates the classification result and the classification result evaluation value. In the observation data classifier, when the classification end instruction is input, the observation data classification in which the best classification result evaluation value is recorded among the observation data classification results obtained by the repetition processing is input to the observation data classifier Output the result to the kind learnerTherefore, when various types are targeted, it is possible to introduce an evaluation standard for classifying to some extent at the classification stage, and there is an effect that efficient learning can be expected.
[0110]
  As described above, the class identification device according to claim 2 of the present invention classifies observation data with a type discriminator that is an observation data with a type discriminator representing a type by the shortest distance method, When the classification number representing the number is reached, an observation data classifier that determines the classification,Observations that have the highest affiliation probability for each classification of observation data m regardless of the presence or absence of a type discriminatorThe maximum membership probability in data m is P_mmaxIf the number of data is Nd, ΣP_mmaxThe classification result evaluator that calculates the classification result evaluation value Ed from / Nd, the classification result evaluation value Ed and a predetermined threshold value are compared, and if the classification result evaluation value Ed is equal to or less than the predetermined threshold value, the classification If it is determined that execution is necessary, assuming that Ncd is the average number of observation data belonging to one class, the number of classification is changed to 1 / Ncd of the number of observation data, and it is determined that classification re-execution is unnecessary. In this case, a classification parameter changer that outputs a classification end instruction and a type of linear discriminant that obtains a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classifier. A classification learner that performs learning, and a type discriminator that discriminates the type of type discrimination target observation data depending on whether it belongs to an area above or below the boundary line for each classification, and the classification parameter Until the reclassifier determines that classification re-execution is unnecessary and outputs the classification end instruction, the observation data classifier outputs the classification and the classification result, and then the classification result evaluator evaluates the classification result and the classification result evaluation value. In the observation data classifier, when the classification end instruction is input, the observation data classification in which the best classification result evaluation value is recorded among the observation data classification results obtained by the repetition processing is input to the observation data classifier Since the results are output to the type learner, there are few observation data with type discriminators, and evaluation is possible even when all observation data is to be classified regardless of the presence or absence of type discriminators. There is an effect that it is possible to obtain a classification reflecting the distribution status of.
[0116]
  The class identification method according to claim 3 of the present invention is as described above.Observation data classifierObservation data classification step that classifies observation data with a type discriminator, which is an observation data with a type discriminator that represents the type, by the shortest distance method, and determines the classification when the number of classifications represents the number of classifications as the convergence condition When,Classification result evaluatorFrom the (number of data of type i belonging to class j) / (total number of data of type i), the membership probability P in class j of type i of the observation data with type discriminator_ijAnd the maximum affiliation probability in type i is P_imaxIf the number of types is Nk, ΣP_imaxA classification result evaluation step of calculating a classification result evaluation value Ek from / Nk;Classification parameter changerWhen the classification result evaluation value Ek is compared with a predetermined threshold, and it is determined that the classification re-execution is necessary when the classification result evaluation value Ek is equal to or less than the predetermined threshold, Ncd belongs to one classification If the average number of observation data is assumed, the classification number is changed to 1 / Ncd of the observation data number and is output. If it is determined that classification re-execution is unnecessary, a classification parameter changing step for outputting a classification end instruction;By type learning device,A type learning step for performing type learning by a linear discriminant method for obtaining a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classification step;By type discriminator,A type determining step for determining the type of the type determining target observation data depending on whether the region belongs to either the upper side or the lower side of the boundary line for each classification. Until the classification end instruction is output, the process of outputting the classification and the classification result in the observation data classification step, and then repeating the process of outputting the evaluation and the classification result evaluation value in the classification result evaluation step is repeated. In the step, when the classification end instruction is input, the observation data classification result in which the best classification result evaluation value is recorded is output to the type learning step among the observation data classification results obtained by the iterative processing. It is said that learning can be performed based on better classification results by evaluating and reclassifying in stages. An effect.
[0117]
  The class identification method according to claim 4 of the present invention is as described above.Observation data classifierObservation data classification step that classifies observation data with a type discriminator, which is an observation data with a type discriminator that represents the type, by the shortest distance method, and determines the classification when the number of classifications represents the number of classifications as the convergence condition When,Observation with classification result evaluator that has the maximum value of belonging to each classification of observation data m regardless of the presence or absence of type discriminatorThe maximum membership probability in data m is P_mmaxIf the number of data is Nd, ΣP_mmaxA classification result evaluation step of calculating a classification result evaluation value Ed from / Nd;Classification parameter changerThe classification result evaluation value Ed is compared with a predetermined threshold, and if the classification result evaluation value Ed is equal to or less than the predetermined threshold and it is determined that classification re-execution is necessary, Ncd belongs to one classification. If the average number of observation data is assumed, the classification number is changed to 1 / Ncd of the observation data number and is output. If it is determined that classification re-execution is unnecessary, a classification parameter changing step for outputting a classification end instruction;By type learning device,A type learning step for performing type learning by a linear discriminant method for obtaining a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classification step;By type discriminator,A type determining step for determining the type of the type determining target observation data depending on whether the region belongs to the upper side or the lower side of the boundary line for each classification, and the classification parameter changing step determines that classification re-execution is unnecessary. Until the classification end instruction is output, the process of outputting the classification and the classification result in the observation data classification step, and then repeating the process of outputting the evaluation and the classification result evaluation value in the classification result evaluation step is repeated. In the step, when the classification end instruction is input, the observation data classification result in which the best classification result evaluation value is recorded is output to the type learning step among the observation data classification results obtained by the iterative processing. It is said that learning can be performed based on better classification results by evaluating and reclassifying in stages. An effect.
[Brief description of the drawings]
FIG. 1 is a diagram showing a configuration of a class identification device according to Embodiment 1 of the present invention.
FIG. 2 is a diagram showing an operation of the class identification device according to Embodiment 1 of the present invention.
FIG. 3 is a diagram showing a configuration of a class identification device according to Embodiment 2 of the present invention.
FIG. 4 is a diagram showing a configuration of a class identification device according to Embodiment 3 of the present invention.
FIG. 5 is a diagram showing an operation of a class identification device according to Embodiment 3 of the present invention.
FIG. 6 is a diagram showing an operation of a class identification device according to Embodiment 3 of the present invention.
FIG. 7 is a diagram showing a configuration of a conventional class identification device.
FIG. 8 is a diagram illustrating an operation of a conventional class identification device.
[Explanation of symbols]
1 observation data classifier, 2 types learner, 3 types discriminator, 4 classification reexecution instruction generator, 41 classification parameter changer, 42 classification method changer, 5 intermediate result evaluator, 51 classification result evaluator, 52 learning Result evaluator, 6A, 6B, 6C class discriminator, 71 learning parameter file, 72 classification parameter file, observation data file with 9 types of discriminator, 10 type discrimination target observation data file, 11 type discrimination result file.

Claims (4)

An observation data classifier that classifies observation data with a type discriminator, which is an observation data with a type discriminator representing a type, by the shortest distance method, and determines the classification when the number of classifications represents the number of classifications as a convergence condition. When,
From the (number of data of type i belonging to class j) / (total number of data of type i), the affiliation probability P _ij in class j of type i of the observation data with type discriminator is calculated, and the maximum affiliation probability in type i And P _imax , and the number of types is Nk, a classification result evaluator that calculates a classification result evaluation value Ek from ΣP _imax / Nk;
When the classification result evaluation value Ek is compared with a predetermined threshold, and it is determined that the classification re-execution is necessary when the classification result evaluation value Ek is equal to or less than the predetermined threshold, Ncd belongs to one classification If the average number of observation data is assumed, the classification number is changed to 1 / Ncd of the number of observation data and output. If it is determined that classification re-execution is unnecessary, a classification parameter changer that outputs a classification end instruction;
A type learner that performs type learning by a linear discriminant that obtains a straight line that is a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classifier;
With a type discriminator that discriminates the type of type discrimination target observation data depending on whether it belongs to the area above or below the boundary line for each classification,
Until the classification parameter changer determines that classification re-execution is unnecessary and outputs the classification end instruction, the observation data classifier outputs the classification and classification result, and then the classification result evaluator evaluates and classifies. Repeat the process of outputting the result evaluation value,
In the observation data classifier, when the classification end instruction is input, the observation data classification result in which the best classification result evaluation value is recorded is output to the type learning device among the observation data classification results obtained by repetition processing. A classifying device characterized by that. An observation data classifier that classifies observation data with a type discriminator, which is an observation data with a type discriminator representing a type, by the shortest distance method, and determines the classification when the number of classifications represents the number of classifications as a convergence condition. When,
Classification result from ΣP _mmax / Nd, where P _mmax is the maximum probability of belonging to observation data m and Nd is the number of data , where the probability of belonging to each classification of observation data m regardless of the presence or absence of a type discriminator is the maximum value. A classification result evaluator for calculating an evaluation value Ed;
The classification result evaluation value Ed is compared with a predetermined threshold, and if the classification result evaluation value Ed is equal to or less than the predetermined threshold and it is determined that classification re-execution is necessary, Ncd belongs to one classification. If the average number of observation data is assumed, the classification number is changed to 1 / Ncd of the number of observation data and output. If it is determined that classification re-execution is unnecessary, a classification parameter changer that outputs a classification end instruction;
A type learner that performs type learning by a linear discriminant that obtains a straight line that is a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classifier;
With a type discriminator that discriminates the type of type discrimination target observation data depending on whether it belongs to the area above or below the boundary line for each classification,
Until the classification parameter changer determines that classification re-execution is unnecessary and outputs the classification end instruction, the observation data classifier outputs the classification and classification result, and then the classification result evaluator evaluates and classifies. Repeat the process of outputting the result evaluation value,
In the observation data classifier, when the classification end instruction is input, the observation data classification result in which the best classification result evaluation value is recorded is output to the type learning device among the observation data classification results obtained by repetition processing. s identification device characterized by. The observation data classifier classifies observation data with a type discriminator, which is an observation data with a type discriminator that represents the type, by the shortest distance method. The observation data classification step to be determined;
The classification result evaluator calculates the affiliation probability P _ij in the classification j of the type i of the observation data with type discriminator from (number of data of type i belonging to classification j) / (total number of data of type i), A classification result evaluation step for calculating a classification result evaluation value Ek from ΣP _imax / Nk, where P _imax is the maximum membership probability in type i and Nk is the number of types;
When the classification parameter changer compares the classification result evaluation value Ek with a predetermined threshold, and when the classification result evaluation value Ek is equal to or less than the predetermined threshold, it is determined that classification re-execution is necessary. If Ncd is the average number of observation data belonging to one class, the number of classifications is changed to 1 / Ncd of the number of observation data, and when it is determined that classification re-execution is unnecessary, the classification for outputting a classification end instruction is output. A parameter change step;
A type learning step for performing type learning by a linear discriminant that obtains a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classification step by a type learner;
The type discriminator includes upper classification for each border, and the type determination step of determining the type of type discrimination object observation data by belongs to which region of the lower,
The classification and classification result are output in the observation data classification step until the classification re-execution is determined to be unnecessary in the classification parameter changing step and the classification end instruction is output, and then the evaluation and classification are performed in the classification result evaluation step. Repeat the process of outputting the result evaluation value,
In the observation data classification step, when the classification end instruction is input, the observation data classification result in which the best classification result evaluation value is recorded is output to the type learning step among the observation data classification results obtained by repetition processing. A class identification method characterized by The observation data classifier classifies observation data with a type discriminator, which is an observation data with a type discriminator that represents the type, by the shortest distance method. The observation data classification step to be determined;
With the classification result evaluator, the maximum probability of belonging to observation data m is P _mmax and the number of data is Nd, where the probability of belonging to each classification of observation data m regardless of whether or not there is a type discriminator is ΣP a classification result evaluation step of calculating a classification result evaluation value Ed from _mmax / Nd;
When the classification parameter changer compares the classification result evaluation value Ed with a predetermined threshold, and when the classification result evaluation value Ed is equal to or less than the predetermined threshold, it is determined that classification re-execution is necessary. If Ncd is the average number of observation data belonging to one class, the number of classifications is changed to 1 / Ncd of the number of observation data, and when it is determined that classification re-execution is unnecessary, the classification for outputting a classification end instruction is output. A parameter change step;
A type learning step for performing type learning by a linear discriminant that obtains a straight line as a boundary based on the distance and direction from all observation data with a type discriminator classified by the observation data classification step by a type learner;
The type discriminator includes upper classification for each border, and the type determination step of determining the type of type discrimination object observation data by belongs to which region of the lower,
The classification and classification result are output in the observation data classification step until the classification re-execution is determined to be unnecessary in the classification parameter changing step and the classification end instruction is output, and then the evaluation and classification are performed in the classification result evaluation step. Repeat the process of outputting the result evaluation value,
In the observation data classification step, when the classification end instruction is input, the observation data classification result in which the best classification result evaluation value is recorded is output to the type learning step among the observation data classification results obtained by repetition processing. A class identification method characterized by

Images