JP4267726B2 - Device for determining relationship between operation signal and operation amount in control device, control device, data generation device, input / output characteristic determination device, and correlation evaluation device - Google Patents

Description

そして、第１分類に属するニューロンモデルは、第１ベクトルＸの成分数（データ要素数）に等しい個数存在し、第２分類に属するニューロンモデルは、第２ベクトルＹの成分数に等しい個数存在する。ここでは、第１ベクトルＸがｎ個の成分を有するものとし、第２ベクトルＹがｍ個の成分を有するものとする。このため、入力層１０には合計（ｎ＋ｍ）個のニューロンモデルが配置されている。
【００３０】
たとえば、二つのデータの相関関係を評価したい場合には、一方のデータを第１ベクトルに対応させ、他方のデータを第２ベクトルに対応させればよい。また、対象装置の入出力特性を決定する場合には、当該対象装置の入力信号を第１ベクトルに対応させ、出力信号を第２ベクトルに対応させればよい。さらに、制御装置における動作信号と操作量との関係を決定する場合には、動作信号を第１ベクトルに対応させ、操作量を第２ベクトルに対応させればよい。
【００３１】
また、第１分類に属するｊ番目のニューロンモデルは、競合層１２に配置されたｉ番目のニューロンモデルと、結合荷重ｗ_ijにてシナプス結合しているものとする。また、第２分類に属するｋ番目のニューロンモデルは、競合層１２に配置されたｉ番目のニューロンモデルと、結合荷重ｕ_ikにてシナプス結合しているものとする。そして、結合荷重ｗ_ijと結合荷重ｕ_ikとにより、競合層１２のｉ番目のニューロンモデルに関する重みベクトルＷ_iが、次式（２）に示すようにして構成される。
【００３２】
【数２】

図２は、以上説明したニューラルネットワークに学習をさせる際の手順を説明するフロー図である。
【００３３】
同フロー図に示すように、上記ニューラルネットワークに学習をさせるには、まずｓ個の学習ベクトルＩ_l（ｌ＝１〜ｓ）を準備する（Ｓ１０１）。この学習ベクトルは、第１ベクトルＸと第２ベクトルＹの対であり、ランダムなベクトルの対であって、それら対の各々対しては個別評価情報も準備されていなければならない。たとえば、本方法により制御装置の特性を決定しようとする場合には、動作信号と操作量の対を学習ベクトルＩ_lに選び、その対が制御装置の入出力とされた場合に制御量がどの様な振る舞いをするかを表す情報等を個別評価情報とすればよい。
【００３４】
次に、学習繰り返し回数Ｉｔｅを０に設定するとともに（Ｓ１０２）、学習ベクトルＩ_lの添え字である変数ｌを０に設定する（Ｓ１０３）。そして、Ｓ１０１で準備した学習ベクトルＩ_lの各成分を、図１に示したニューラルネットワークの入力層１０に配置されたニューロンモデルに入力する（Ｓ１０４）。
【００３５】
入力層１０に学習ベクトルＩ_１が与えられると、勝者ユニット（勝ちニューロン）、即ち学習ベクトルＩ_１に対して最も反応する競合層１２のニューロンモデルが決定される（Ｓ１０５）。具体的には、次式（３）により、競合層１２のｉ番目のニューロンモデルと入力層１０のニューロンモデルとの間のユークリッド距離Ｄ _ｉを求め、それが最も短いものを勝者ユニットとする。
【００３６】
【数３】

なお、勝者ユニットを決定する場合、次式（４）に示すファジイ類似性尺度Ｓ_iを競合層１２の各ニューロンモデルについて算出し、それが最も小さな値をとるニューロンモデルを勝者ユニットとしてもよい。同式（４）において、ｘ∧ｗは、ｘとｗのいずれか小さな値を意味する。またｘ∨ｗは、ｘとｗのいずれか大きな値を意味する。
【００３７】
【数４】

次に、本方法では、勝者ユニットの近傍に位置するニューロンモデル、即ち近傍ユニットを決定する（Ｓ１０６）。たとえば、競合層１２のｉ番目のニューロンモデルを勝者ユニットとした場合、競合層１２のニューロンモデルが図１に示すように直線状に配置されていると仮定すれば、競合層（ｉ−２）番目から（ｉ＋２）番目のニューロンモデル等、前後するニューロンモデルを近傍ユニットに選べばよい。
【００３８】
その後、本方法では、勝者ユニットと近傍ユニットについて、重みベクトルＷ_iを次式（５）及び（６）により更新する（Ｓ１０７）。
【００３９】
【数５】

まず、上式（５）は学習ベクトルＩ_lに対する評価が肯定的である場合に用いられる更新式である。同式（５）においてＷ_i ^oldは更新前の重みベクトルＷ_iであり、Ｗ_i ^newは更新後の重みベクトルＷ_iである。また、α（ｔ）は学習係数であり、時間とともに減少するｔの関数が採用されている。さらに、Ｅ_lは正の実数値をとる評価値であり、学習ベクトルＩ_lに与えられる肯定的な個別評価情報を表す。
【００４０】
この式（５）による重みベクトルＷ_iの更新により、学習ベクトルＩ_lに対する評価が肯定的であれば、勝者ユニット及びその近傍ユニットの重みベクトルＷ_iは、学習係数α（ｔ）及び評価値Ｅ_lに比例して、学習ベクトルＩ_lに近づくことになる。
【００４１】
これに対して上式（６）は学習ベクトルＩ_lに対する評価が否定的である場合に用いられる更新式である。同式（６）において、ｓｉｇｎは符号関数である。また、Ｅ_lは負の実数値をとる評価値であり、学習ベクトルＩ_lに与えられる否定的な個別評価情報を表す。
【００４２】
この式（６）による重みベクトルＷ_iの更新により、学習ベクトルＩ_lに対する評価が否定的であれば、勝者ユニット及びその近傍ユニットの重みベクトルＷ_iは学習ベクトルＩ_lから遠ざかることになる。その際、重みベクトルＷ_iが学習ベクトルＩ_lから遠ざかる距離は学習係数α（ｔ）及び評価値Ｅ_lに比例する。また、同式（６）の右辺の指数関数の働きにより、より近い重みベクトルＷ_jがより大きく学習ベクトルＩ_lから遠ざけられることになる。
【００４３】
その後、変数ｌが最大値ｓ未満であれば（Ｓ１０８）、ｌをインクリメントし（Ｓ１１０）、次の学習ベクトルＩ_lをニューラルネットワークの入力層１０に与える（Ｓ１０４）。一方、変数ｌが最大値ｓに達すれば（Ｓ１０８）、繰り返し回数Ｉｔｅが設定値ＩＴＥに達しているかを判断し（Ｓ１０９）、達していなければ繰り返し回数Ｉｔｅをインクリメントするとともに（Ｓ１１１）、変数ｌを０に戻し（Ｓ１０３）、再び学習ベクトルＩ０をニューラルネットワークの入力層１０に与える（Ｓ１０４）。
【００４４】
以上のように、本方法ではニューラルネットワークの学習ベクトルとして、対象装置の入出力データ等の二つのデータの対を用いている。こうしてニューラルネットワークに学習をさせると、以下に説明するように、ニューラルネットワークの結合荷重から、ある評価を与えられるデータ対の相関関係が得られるようになる。
【００４５】
図３及び図４は、第１ベクトルＸ及び第２ベクトルＹとしていずれも１成分のみを有するものを採用した場合の、ニューラルネットワークの学習の様子を模式的に示す図である。この場合、学習ベクトルＩは２成分を有し、重みベクトルＷ_iは第１ベクトルＸ及び第２ベクトルＹに対応して２成分を有することになる。
【００４６】
図３には、上から学習ベクトル空間（平面）、重みベクトル空間（平面）、競合層１２が順に描かれている。学習ベクトル空間は第１ベクトルＸ及び第２ベクトルＹに対応する座標軸を有しており、学習ベクトルＩが与えられると、それを平面上の一点に対応づけることができるようになっている。一方、重みベクトル空間は重みベクトルＷの２成分に対応する軸を有しており、重みベクトルＷ_ｉの終点が平面上の一点にそれぞれ描かれるようになっている。
【００４７】
同図に示すように、ある学習ベクトルＩが与えられると、競合層１２に配置されるニューロンモデルから勝者ユニットが決定される。この勝者ユニットは、学習ベクトルＩを重みベクトル空間上の一点に表した場合に、その点に最も近い点に終点を有する重みベクトルＷ_ｉである。続いて、その勝者ユニットに対する近傍ユニットも選ばれる。同図では競合層１２にニューロンモデルを平面的に配置しているため、勝者ユニットの周囲を取り巻くユニットが近傍ユニットに選ばれる。
【００４８】
そして、与えられた学習ベクトルＩに対する個別評価が肯定的であって評価値Ｅが正の値をとる場合、それら勝者ユニット及び近傍ユニットの重みベクトルＷ_iが上式（５）に従って更新される。この場合、近傍ユニットの重みベクトルＷ_iは必ずしも勝者ユニットの重みベクトルＷ_iに類似するものとはなっていない。しかしながら、学習が進むに従って勝者ユニットの重みベクトルＷ_iと近傍ユニットの重みベクトルＷ_iとは次第に類似するようになる。
【００４９】
なお、同図では特に示さないが、入力される学習ベクトルＩに対する評価が否定的である場合、即ち評価値Ｅが負の値をとる場合、上式（６）に従って、勝者ユニット及び近傍ユニットの重みベクトルＷ_iは当該学習ベクトルＩから遠ざかるように更新される。
【００５０】
図４（ａ）は、学習が進んだ後の重みベクトル空間の様子を示す図である。同図に示すように、学習が相当程度に進むと重みベクトルＷ_iの終点を表す各点は肯定的な評価を与えられた学習ベクトルＩの近傍に集結する。従って、この重みベクトルＷ_iの分布を調べることにより、学習ベクトルＩに含まれる第１ベクトルＸと第２ベクトルＹとの関係を評価することができる。
【００５１】
たとえば、第１ベクトルＸとして制御装置に入力される動作信号を採用し、第２ベクトルＹとして制御装置から出力される操作量を採用した場合、それら動作信号及び操作量のデータ対と個別評価とを予め十分にニューラルネットワークに学習させると、同図（ａ）に示すように、安定的に制御量を制御できる動作信号と操作量とのデータ対が重みベクトル空間のおける多数の重みベクトルＷ_iの終点として得られることになる。このため、例えば同図（ａ）に示される重みベクトルＷ_iの終点を補間する等して、同図（ｂ）に示すような、動作信号と操作量との関係を表す関数を容易に得ることができる。この結果、その関数を制御装置に予め設定することにより、ある動作信号を与えられると適切な操作量を出力する制御装置を容易に設計することができる。この点は、対象装置の入出力特性を決定する場合にも同様に当てはまる。
【００５２】
また、以上のように、第１ベクトルＸと第２ベクトルＹの相関関係を、ニューラルネットワークの学習後に一旦評価してもよいが、そのような評価をせずとも、第１ベクトルＸに対応する好ましい（肯定的評価が得られるであろう）第２ベクトルＹを、重みベクトルＷ_iに基づいて算出することができる。
【００５３】
図５は、重みベクトルＷ_iから第１ベクトルＸに対応する好ましい第２ベクトルＹを算出する手順を説明するフロー図である。
【００５４】
この手順では、図６に示すようにして、まず第１ベクトルＸのみをニューラルネットワークの入力層１０に与える。ここでの第１ベクトルＸは、学習段階で用いた学習ベクトルの構成要素たる第１ベクトルＸと区別するため、特に事実入力ベクトルＸ⁰と記す（Ｓ２０１）。また以下では、この事実入力ベクトルＸ⁰に対応する第２ベクトルＹを、学習段階で用いた学習ベクトルの構成要素たる第２ベクトルＹと区別するため、特に事実出力ベクトルＹ⁰と記す。
【００５５】
次に、事実入力ベクトルＸ⁰に対する競合層１２上の各ニューロンモデルの出力ｚ_p（ここでは、「部分反応度」という。）を、次式（７）に従って算出する（Ｓ２０２）。同式（７）においてβは所与の定数である。
【００５６】
【数６】

その後、Ｓ２０２で得られた部分反応度ｚ_pと結合荷重Ｕ_prとに基づき、次式（８）に従って事実出力ベクトルＹ⁰を算出する（Ｓ２０３）。図７は、この事実出力ベクトルＹ⁰の算出処理の概念を示すものである。
【００５７】
【数７】

同式（８）は結合荷重Ｕ_prを部分反応度ｚ_pを用いて重み付け加算すると、事実出力ベクトルＹ⁰のｒ番目の成分が算出できることを表している。
【００５８】
たとえば、第１ベクトルＸ及び第２ベクトルＹとしていずれも１成分のみを有するものを採用した場合、図８に示すようにして重みベクトルＷが重みベクトル空間上に表されていれば、ある事実入力ベクトルＸ⁰が与えられると、以上説明した処理では、領域１４内に重みベクトルＷ_iの終点を有するニューロンモデルを中心として、式（７）で算出される部分反応度ｚ_pが大きな値を有することになる。そして、式（８）を計算すると、この領域１４内に重みベクトルＷ_iの終点を有するニューロンモデルについての結合荷重ｕの値が支配的となり、同図の矢印１６の辺りに事実出力ベクトルＹ⁰の値が得られることになる。
【００５９】
以上のようにして、重みベクトルＷ_iから第１ベクトルＸに対応する好ましい第２ベクトルＹを算出することができる。
【００６０】
実施の形態２．
実施の形態２では、上述した実施の形態１に係る方法を応用した適応制御装置について説明する。
【００６１】
図９は、本実施の形態に係る適応制御装置の構成を示す図である。同図に示す適応制御装置は、図１に示したニューラルネットワークをハードウェアとして維持・更新するものであり、制御部４４と学習部４６との間で第１結合荷重記憶部２０及び第２結合荷重記憶部２２が共有されており、双方から随時アクセスすることができるようになっている。この第１結合荷重記憶部２０は図１に示したニューラルネットワークにおける結合荷重ｗ_ijを記憶するものであり、第２結合荷重記憶部は図１に示したニューラルネットワークにおける結合荷重ｕ_ikを記憶するものである。
【００６２】
まず、学習部４６は、結合荷重更新部２８と、評価部３８と、第１及び第２結合荷重記憶部２０，２２と、を含んでいる。評価部３８は、制御対象４０から得られる制御量を評価して、評価値Ｅを結合荷重更新部２８に与えるようになっている。評価値Ｅには、具体的には制御量の安定性等を表す情報を採用することができ、その値は−１〜＋１に亘って連続する値をとる。
【００６３】
結合荷重更新部２８には、制御部４４から出力される操作量Ｙもまたスイッチ３４及び遅延素子３２を介して入力されるようになっている。また、基準入力とフィードバック量との差分である動作信号Ｘも、加算器４２により生成され、遅延素子３０を介して結合荷重更新部２８に入力されるようになっている。なお、遅延素子３０，３２は、動作信号Ｘと操作量Ｙと評価値Ｅとが互いに対応するものとなるように、遅延時間がそれぞれ設定される。また、動作信号Ｘはｎ個のデータ要素を有し、操作量Ｙはｍ個のデータ要素を有し、それらは複数本の信号線により伝送される。
【００６４】
そして、結合荷重更新部２８では、上式（５）及び（６）に基づいて重みベクトルＷ_iをリアルタイムに更新し、更新後の結合荷重ｗ_ij，ｕ_ikを第１及び第２結合荷重記憶部２０，２２にそれぞれ格納する。
【００６５】
一方、制御部４４は、部分反応度算出部２４と、操作量算出部２６と、第１及び第２結合荷重記憶部２０，２２と、を含んでいる。部分反応度算出部２４には第１結合荷重記憶部２０から結合荷重ｗ_pqが入力されるとともに、加算器４２から事実動作信号（事実入力ベクトル）Ｘ⁰が入力されるようになっている。そして、部分反応度算出部２４では、これらの情報をもとに、上式（７）に従って部分反応度ｚ_pを全ての競合層１２上のニューロンモデルについて算出する。
【００６６】
また、操作量算出部２６には第２結合荷重記憶部２２から結合荷重ｕ_prが入力されるとともに、部分反応度算出部２４から部分反応度ｚ_pが入力されるようになっている。そして、操作量算出部では、これらの情報をもとに、上記式（８）に従って事実操作量（事実出力ベクトル）Ｙ⁰を算出する。そして、算出された事実操作量Ｙ⁰が制御対象４０に与えられる。
【００６７】
以上説明した適応制御装置では、制御対象４０の安定性等の諸評価要素に基づいて第１及び第２結合荷重記憶部２０，２２の内容が随時更新される。そして、その随時更新される第１及び第２結合荷重記憶部２０，２２の内容に従って、事実動作信号Ｘ⁰に対応する、好ましい（肯定的な評価を得られるであろう）事実操作量Ｙ⁰が制御部４４により算出され、制御対象４０に与えられる。このため、制御対象４０の特性が時とともに変化した場合であっても、その変化に適応的に追従し、好適な制御特性を維持することができる。
【００６８】
なお、以上説明した本発明の実施の形態１及び２は種々の変形実施が可能である。たとえば、以上の説明ではニューラルネットワークの結合荷重ｗ_ij，ｕ_ikの更新、即ち重みベクトルＷ_iの更新を、式（５）及び式（６）に従って行うようにしたが、その他の更新方法を採用してもよい。
【００６９】
また、評価値Ｅ_lやＥとしては、学習ベクトルに対応してあらゆる情報を採用することができる。たとえば、学習ベクトル毎の確信度、重要度、刺激程度などを評価値Ｅ_lやＥとして採用することができる。
【００７０】
さらに、実施の形態２に係る適応制御装置では、基準入力としてランダムな値を加算器４２に与えるとともに、スイッチ３４を端子３６側に切り換えて該端子３６からランダムな値を与え、実際の装置の使用に先立ち、ニューラルネットワークの学習をある程度済ませておくようにしてもよい。
【図面の簡単な説明】
【図１】 本発明の実施の形態１に係る方法で用いるニューラルネットワークの構成を示す図である。
【図２】 図１に示すニューラルネットワークの学習手順を説明するフロー図である。
【図３】 図１に示すニューラルネットワークの学習の様子を説明する図である。
【図４】 学習後の重みベクトル空間の様子を示す図である。
【図５】 図１に示すニューラルネットワークを用い、事実入力ベクトルから事実出力ベクトルを算出する手順を説明するフロー図である。
【図６】 図１に示すニューラルネットワークを用いて、事実入力ベクトルに対する競合層の各ニューロンモデルの部分反応度が算出される様子を示す図である。
【図７】 競合層の各ニューロンモデルの部分反応度に基づいて、事実入力ベクトルに対応する事実出力ベクトルが算出される様子を示す図である。
【図８】 事実出力ベクトルの計算過程を説明する図である。
【図９】 本発明の実施の形態２に係る適応制御装置の構成を示す図である。
【符号の説明】
１０ 入力層、１２ 競合層、２０ 第１結合荷重記憶部、２２ 第２結合荷重記憶部、２４ 部分反応度算出部、２６ 操作量算出部、２８ 結合荷重更新部、３０，３２ 遅延素子、３４ スイッチ、３６ 端子、３８ 評価部、４０制御対象、４２ 加算器、４４ 制御部、４６ 学習部。[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a control device manufacturing method, a control device, a data generation device, an input / output characteristic determination method, and a correlation evaluation method, and more particularly to a control device manufacturing method, a control device, a data generation device, an input using a neural network. The present invention relates to an output characteristic determination method and a correlation evaluation method.
[0002]
[Prior art and problems to be solved by the invention]
In designing a high-performance control device, accurate information about the characteristics of the controlled object and the environment in which the controlled object is placed must be obtained in advance. However, in reality, it is often difficult to obtain such information correctly.
[0003]
For this reason, the characteristics of the control device are determined by actually giving a large number of pairs of operation signals and manipulated variables to the object to be controlled and investigating what control results are obtained. It is practical.
[0004]
At this time, if a method for evaluating the relation between the operation signal and the operation amount is proposed based on a large number of pairs of the operation signal and the operation amount and the individual evaluation on the pair, a preferable operation signal is obtained based on the evaluation result. It is possible to easily design and manufacture the control device by determining the relationship with the operation amount.
[0005]
In addition to the characteristics of the control device as described above, generally, if a method for evaluating the correlation between two data (including vector data) based on individual evaluation for each data pair is proposed. Is meaningful.
[0006]
In this regard, neural network technology is the best technology for processing individual data and individual evaluations. However, existing neural network models simply classify input data or simply map data. No neural network model that can evaluate the correlation between two data has been proposed yet.
[0007]
The present invention has been made in view of the above problems, and a first object is to control the characteristics of a control device that can be easily determined based on individual evaluation of a plurality of pairs of operation signals and operation amounts. It is to provide a method for manufacturing an apparatus.
[0008]
A second object is to provide a control device that allows a neural network to learn preferable characteristics of the control device and control a controlled object in accordance with the learning content.
[0009]
A third object is to provide a data generation device that can cause a neural network to learn the correlation between two data and the evaluation thereof, and generate data corresponding to certain data according to the learning content. There is.
[0010]
A fourth object is to provide an input / output characteristic determination method capable of easily determining the input / output characteristics of a target device based on individual evaluation for a plurality of pairs of input / output data.
[0011]
Furthermore, the fifth object is to provide a correlation evaluation method capable of evaluating the correlation between two data based on the individual evaluation for each data pair.
[0012]
[Means for Solving the Problems]
(1) In order to solve the above problems, a control device according to the present invention.For determining relationship between operation signal and operation amountIsPlaced in the first layer of the neural network,A neuron model for motion signals corresponding to individual data elements constituting the motion signal, and a neuron model for operation amounts corresponding to the individual data elements constituting the operation amount,A first layer neuron model group including,Arranged in the second layer of the neural network,Combines with the neuron model for motion signals and manipulated variablesSecond layerNeuron modelGroup andThe firstStratumOolong modelgroupAre sequentially input a pair of learning operation signal and operation amount, and the pair of operation signal and operation amount and the corresponding operation signal and operation amount pair.Includes an evaluation value representing the degree of positive or negative evaluationBased on the individual evaluation information, the neuron model for the operation signal and the manipulated variable and the second modelStratumUpdated the coupling load with the uron model groupLearning means for performing at least one neuron model in the second-layer neuron model group on a connection load between the neuron model and each neuron model included in the first-layer neuron model group. If the individual evaluation information includes a positive evaluation value with respect to a learning vector having each element as an individual data element constituting the learning operation signal and operation amount, the evaluation value When the individual evaluation information includes a negative evaluation value, the calculation is performed to move away by a distance proportional to the evaluation value, and each element of the weight vector as a result of the calculation is associated with the corresponding combined load. The learning means to be the updated value ofThe input / output characteristics of the control device are determined based on the updated combined loadA determination unit for determining a connection load between each neuron model included in the second-layer neuron model group and the neuron model for motion signal as an input to the control device, and including the second-layer neuron model group in the second-layer neuron model group When the connection load between each neuron model to be operated and the neuron model for operation signals is the output of the control device, an operation is performed to obtain a function representing the relationship between the input and the output, and the result of this operation And determining means for setting an input / output characteristic of the control device as a function ofIt is characterized by that.
[0013]
  In the present invention, a pair of learning motion signals and manipulated variables is sequentially input to the first layer neuron model, and the individual evaluation information such as the learning motion signal and manipulated variable pairs and the stability of the control amount, for example. Based on the above, the combined load is updated. By doing this, the connection weights are sequentially updated so that the neuron model located nearby in the second layer reacts to a pair of operation signals and manipulated variables with similar input / output relationships. Based on the combined load, a pair of an operation signal and an operation amount corresponding to a certain individual evaluation can be obtained. As a result, for example, a pair of an operation signal and an operation amount that stabilizes the control amount can be obtained from the combined load, and the characteristics of the control device can be easily determined.Dobe able to.
[0014]
(2) Further, the control device according to the present invention is a control device using a neural network, and is an operation signal that is arranged in the first layer of the neural network and receives individual data elements constituting the operation signal. A first layer neuron model group including a neuron model for operation, and a neuron model for operation amount to which individual data elements constituting the operation amount are input, and arranged on the second layer of the neural network for the operation signal And a second layer neuron model group coupled to the operation amount neuron model, an operation signal and an operation amount input to the first layer neuron model group, and the operation signal and the operation amountIncludes an evaluation value representing the degree of positive or negative evaluationLearning means for updating a connection load between the first layer neuron model group and the second layer neuron model group based on the individual evaluation informationA weight having at least one neuron model of the second-layer neuron model group as a component with a connection weight between the neuron model and each neuron model included in the first-layer neuron model group. When the individual evaluation information includes a positive evaluation value, the vector is brought closer to a distance proportional to the evaluation value with respect to the learning vector having each data element constituting the operation signal and the operation amount as the element. When the individual evaluation information includes a negative evaluation value, the calculation is performed to move away by a distance proportional to the evaluation value, and each element of the weight vector as a result of this calculation Learning meansWhen,An operation signal is input to the operation signal neuron model,The motion signal and the combined loadBased on the product of the value obtained by substituting the value into a preset equation and the neuron model included in the second layer neuron model group and the connection weight between the neuron models for operation The valueOperation amount to be output in response to the operation signalAsAnd an operation amount calculating means for calculating.
[0015]
As described above, information about a pair of an operation signal and an operation amount whose individual evaluation is positive is stored in the combined load. Therefore, in the present invention, when a certain motion signal is given, an operation amount corresponding to the motion signal can be calculated based on the motion signal and the coupling load. For example, when an operation signal is given, the operation amount calculation means can calculate a corresponding operation amount and a high stability of the control amount based on the operation signal and the coupling load. For this reason, according to the control device according to the present invention, it is possible to cause the neural network to learn preferable characteristics of the control device, and to control the control target according to the learning content. Further, if learning by the learning means is performed in parallel with the control of the control target, the control target can be adaptively controlled.
[0016]
  In one aspect of the present invention, the operation amount calculation meansThe preset formula in is, Included in the second layer neuron model groupOneConnection load between neuron model and neuron model for motion signalVector with each elementAnd operation signalIs a formula that takes a larger value as the distance between the vector and each vector constituting each data element is smaller, and the manipulated variable calculation meansEach neuron model included in the second layer neuron model groupThe calculation is performed by substituting the connection load between the neuron model and the neuron model for motion signal and the motion signal input to the neuron model for motion signal into the preset equation. The result value is assigned to each neuron model included in the second-layer neuron model group.Partial reactivity of the motion signal toAsMeans for calculating; a neuron model included in the second layer neuron model group; andoperationThe connection load between the neuron model for quantity andThe neuron models included in the two-layer neuron model groupThe partial reactivity;Product ofBased onThe valueManipulation amountAsAnd means for calculating.
[0017]
  In this aspect, each partial reactivity of the second layer neuron model group with respect to the input of the motion signal is first calculated. Then, based on the partial reactivity and the connection load between the control neuron model and the second layer neuron model group, it corresponds to the operation signal.operationCalculate the amount. For example, the connection load between the control neuron model and the second layer neuron model group is weighted and added by the partial reactivity,operationWhat is necessary is just to calculate quantity. In this way, based on the learning contents of the neural network, it corresponds to the operation signal.operationThe amount can be calculated.
[0018]
(3) A data generation device according to the present invention is a data generation device that generates second data based on first data, and is arranged in a first layer of a neural network, and each of the first data A first layer neuron model group including a first neuron model corresponding to a data element and a second neuron model corresponding to an individual data element of the second data; and a second layer of the neural network, A second layer neuron model group coupled to the first and second neuron models;Inputting first data into the first neuron model;A connection load between the first layer neuron model group and the second layer neuron model group;A value based on a product of a value obtained by performing the calculation by substituting the first data into a preset equation and a connection weight between the second layer neuron model group and the second neuron model TheSecond dataAsIn the learning stage, a pair of first and second data for learning is sequentially input to the first layer neuron model group, and the pair of first and second data for learning is included. And against itIncludes an evaluation value representing the degree of positive or negative evaluationBased on the individual evaluation information, the connection weight between the first layer neuron model group and the second layer neuron model group is updated.In this update process, for at least one neuron model in the second layer neuron model group, a connection load between the neuron model and each neuron model included in the first layer neuron model group is set as each element. When the individual evaluation information includes a positive evaluation value with respect to a learning vector having individual data elements of the first and second data for learning as each element, the weight vector to be proportional to the evaluation value When the individual evaluation information includes a negative evaluation value, the calculation is performed to move away by a distance proportional to the evaluation value, and each element of the weight vector as a result of this calculation is updated for the corresponding combined load. Later valueIt is characterized by that.
[0019]
According to the present invention, the neural network learns the correlation between the first data and the second data and the evaluation thereof, and generates the second data corresponding to the given first data according to the learning content. Can do.
[0020]
(4) Also, input / output characteristics determination according to the present inventionapparatusDetermines the input / output characteristics that determine the relationship between the input data and output data of the target device.apparatusBecausePlaced in the first layer of the neural network,A neuron model for input data corresponding to each data element of the input data, and a neuron model for output data corresponding to each data element of the output data,Include1st layerNeuron model group and,Arranged in the second layer of the neural network,Combine with the neuron model for input data and output dataSecond layerNeuron model groupWhen,The firstStratumOolong modelgroupSequentially input a pair of input and output data for learning, and the pair of data and corresponding dataIncludes an evaluation value representing the degree of positive or negative evaluationBased on the individual evaluation information, the input data and output data neuron models and the second dataStratumUpdated the coupling load with the uron model groupLearning means for performing at least one neuron model in the second-layer neuron model group on a connection load between the neuron model and each neuron model included in the first-layer neuron model group. When the individual evaluation information includes a positive evaluation value with respect to a learning vector having individual elements of the input data and output data as the weight vector, the distance is proportional to the evaluation value. When the individual evaluation information includes a negative evaluation value, the calculation is performed to move away by a distance proportional to the evaluation value, and each element of the weight vector as a result of this calculation is the updated value of the associated load And learning means and beforeDetermine the input / output characteristics of the target device based on the combined loadA determining unit, wherein a connection load between each neuron model included in the second layer neuron model group and the neuron model for input data is set as an input of the target device, and is included in the second layer neuron model group; When the connection load between each neuron model and the output data neuron model is the output of the target device, an operation is performed to obtain a function representing the relationship between the input and the output. Determining means for setting a function as an input / output characteristic of the target device.It is characterized by that.
[0021]
In this way, the input / output data of the target device and its individual evaluation can be learned by the neural network, and the input / output characteristics of the target device can be determined based on the learning result.
[0022]
(5) Correlation evaluation according to the present inventionapparatusIs a correlation evaluation that evaluates the correlation between the first data and the second dataapparatusBecausePlaced in the first layer of the neural network,A first data neuron model corresponding to each data element of the first data, and a second data neuron model corresponding to each data element of the second data.A first layer neuron model group including,Arranged in the second layer of the neural network,Combine with the first and second data neuron modelsSecond layerNeuron model groupWhen,The firstStratumOolong modelgroupSequentially input a pair of first and second data for learning, and the pair of data and the corresponding data pairIncludes an evaluation value representing the degree of positive or negative evaluationBased on the individual evaluation information, the first and second data neuron models and the second dataStratumUpdated the coupling load with the uron model groupLearning means for performing at least one neuron model in the second-layer neuron model group on a connection load between the neuron model and each neuron model included in the first-layer neuron model group. When the individual evaluation information includes a positive evaluation value with respect to a learning vector in which the individual data elements of the first and second data for learning are each element, When the individual evaluation information includes a negative evaluation value, the calculation is performed to move away by a proportional distance, and to move away by a distance proportional to the evaluation value. Learning means to be the updated value,Evaluating the correlation between the first data and the second data based on the bond weightEvaluation means, wherein a connection load between each neuron model included in the second layer neuron model group and the first data neuron model is defined as first data, and each of the second layer neuron model groups includes When the connection load between the neuron model and the second data neuron model is the second data, an operation is performed to obtain a function representing the relationship between the first data and the second data. Evaluation means, wherein the resulting function is a function representing a correlation between the first data and the second data.It is characterized by that.
[0023]
In this way, the correlation between the two data can be evaluated based on the individual evaluation for each data pair.
[0024]
DETAILED DESCRIPTION OF THE INVENTION
DESCRIPTION OF EXEMPLARY EMBODIMENTS Hereinafter, preferred embodiments of the invention will be described in detail with reference to the drawings.
[0025]
Embodiment 1 FIG.
In the first embodiment, using a neural network, a method for evaluating the correlation between two data and a method for determining input / output characteristics of a target device, in particular, a relationship between an operation signal and an operation amount in a control device is determined. A method will be described.
[0026]
In the following description, alphabetic characters representing various variables are used, but the uppercase alphabet represents a vector quantity and the lowercase alphabet represents a scalar quantity.
[0027]
FIG. 1 is a diagram for explaining a method according to this embodiment, and shows a configuration of a neural network used in this method. In the neural network shown in the figure, a learning vector I is input to a neuron model (hereinafter also referred to as “unit”) arranged in the input layer 10, and the neuron model of these input layers 10 is arranged in the competitive layer 12. Synapse with the neuron model.
[0028]
The neuron models arranged in the input layer 10 are roughly classified into two. The neuron model to which the first vector X is input belongs to the first category, and the neuron model to which the second vector Y is input belongs to the second category. That is, the learning vector I is composed of the first vector X and the second vector Y as shown in the following equation (1).
[0029]
[Expression 1]

The number of neuron models belonging to the first class is equal to the number of components (data elements) of the first vector X, and the number of neuron models belonging to the second class is equal to the number of components of the second vector Y. . Here, it is assumed that the first vector X has n components, and the second vector Y has m components. For this reason, a total of (n + m) neuron models are arranged in the input layer 10.
[0030]
For example, when it is desired to evaluate the correlation between two data, one data may be associated with the first vector and the other data may be associated with the second vector. Further, when determining the input / output characteristics of the target device, the input signal of the target device may correspond to the first vector, and the output signal may correspond to the second vector. Furthermore, when determining the relationship between the operation signal and the operation amount in the control device, the operation signal may correspond to the first vector and the operation amount may correspond to the second vector.
[0031]
Further, the j-th neuron model belonging to the first classification is the same as the i-th neuron model arranged in the competitive layer 12 and the connection weight w._ijIt is assumed that the synapse is connected. In addition, the kth neuron model belonging to the second category is different from the ith neuron model arranged in the competitive layer 12 with the connection weight u._ikIt is assumed that the synapse is connected. And the combined load w_ijAnd combined load u_ikAnd the weight vector W for the i-th neuron model of the competitive layer 12_iIs configured as shown in the following equation (2).
[0032]
[Expression 2]

FIG. 2 is a flowchart for explaining a procedure when the above-described neural network is trained.
[0033]
As shown in the flowchart, in order to make the neural network learn, first, s learning vectors I_l(L = 1 to s) is prepared (S101). This learning vector is a pair of the first vector X and the second vector Y, and is a pair of random vectors, and individual evaluation information must be prepared for each of these pairs. For example, when determining the characteristic of the control device by this method, the pair of the operation signal and the manipulated variable is expressed as a learning vector I._lThe information indicating how the control amount behaves when the pair is input / output of the control device may be used as the individual evaluation information.
[0034]
Next, the learning repetition number Ite is set to 0 (S102), and the learning vector I_lIs set to 0 (S103). And the learning vector I prepared in S101_lAre input to the neuron model arranged in the input layer 10 of the neural network shown in FIG. 1 (S104).
[0035]
  Learning vector I in the input layer 10₁Is given, the winner unit (winning neuron), ie the learning vector I₁The neuron model of the competitive layer 12 that most responds to is determined (S105). Specifically, the Euclidean distance between the i-th neuron model of the competitive layer 12 and the neuron model of the input layer 10 is expressed by the following equation (3).D _iAnd the shortest one is the winner unit.
[0036]
[Equation 3]

When determining the winner unit, the fuzzy similarity measure S shown in the following equation (4):_iMay be calculated for each neuron model of the competitive layer 12, and the neuron model having the smallest value may be determined as the winner unit. In the equation (4), x∧w means a smaller value of either x or w. X∨w means a larger value of x and w.
[0037]
[Expression 4]

Next, in this method, a neuron model located in the vicinity of the winner unit, that is, a neighboring unit is determined (S106). For example, when the i-th neuron model of the competitive layer 12 is a winner unit, assuming that the neuron model of the competitive layer 12 is linearly arranged as shown in FIG. 1, the competitive layer (i-2) The preceding and following neuron models such as the (i + 2) -th neuron model may be selected as the neighborhood unit.
[0038]
Thereafter, the method uses the weight vector W for the winner unit and neighboring units._iIs updated by the following equations (5) and (6) (S107).
[0039]
[Equation 5]

First, the above equation (5) represents the learning vector I_lThis is an update formula used when the evaluation of is positive. In equation (5), W_i ^oldIs the weight vector W before update_iAnd W_i ^newIs the updated weight vector W_iIt is. Α (t) is a learning coefficient, and a function of t that decreases with time is adopted. In addition, E_lIs an evaluation value taking a positive real value, and the learning vector I_lRepresents positive individual evaluation information given to.
[0040]
Weight vector W by this equation (5)_iBy updating the learning vector I_lIs positive, the weight vector W of the winner unit and its neighboring units_iIs the learning coefficient α (t) and the evaluation value E_lIs proportional to the learning vector I_lWill approach.
[0041]
On the other hand, the above equation (6) represents the learning vector I_lThis is an update formula used when the evaluation of is negative. In the equation (6), sign is a sign function. E_lIs an evaluation value taking a negative real value, and the learning vector I_lRepresents negative individual evaluation information given to.
[0042]
Weight vector W by this equation (6)_iBy updating the learning vector I_lIs negative, the weight vector W of the winner unit and its neighboring units_iIs the learning vector I_lIt will be away from. At that time, the weight vector W_iIs the learning vector I_lThe distance away from the learning coefficient α (t) and the evaluation value E_lIs proportional to Further, a closer weight vector W is obtained by the action of the exponential function on the right side of the equation (6)_jIs larger learning vector I_lIt will be kept away from.
[0043]
Thereafter, if the variable l is less than the maximum value s (S108), l is incremented (S110), and the next learning vector I_lIs given to the input layer 10 of the neural network (S104). On the other hand, if the variable l reaches the maximum value s (S108), it is determined whether the iteration number It has reached the set value ITE (S109). If not, the iteration number It is incremented (S111), and the variable l Is returned to 0 (S103), and the learning vector I0 is again applied to the input layer 10 of the neural network (S104).
[0044]
As described above, this method uses a pair of two data such as input / output data of the target device as a learning vector of the neural network. When the neural network is trained in this way, as will be described below, the correlation between data pairs given a certain evaluation can be obtained from the coupling load of the neural network.
[0045]
3 and 4 are diagrams schematically illustrating how the neural network learns when both the first vector X and the second vector Y have only one component. In this case, the learning vector I has two components and the weight vector W_iHas two components corresponding to the first vector X and the second vector Y.
[0046]
  In FIG. 3, the learning vector space (plane), the weight vector space (plane), and the competitive layer 12 are drawn in order from the top. The learning vector space has coordinate axes corresponding to the first vector X and the second vector Y. When a learning vector I is given, it can be associated with one point on the plane. On the other hand, the weight vector space isweightHas axes corresponding to the two components of the vector W;weightVector W_iThe end point is drawn at one point on the plane.
[0047]
  As shown in the figure, when a certain learning vector I is given, a winner unit is determined from a neuron model arranged in the competitive layer 12. This winner unit has an end point at a point closest to that point when the learning vector I is represented as one point on the weight vector space.weightVector W_iIt is. Subsequently, a neighboring unit for the winner unit is also selected. In the figure, since the neuron model is arranged in a plane in the competitive layer 12, a unit surrounding the winner unit is selected as a neighboring unit.
[0048]
If the individual evaluation for the given learning vector I is affirmative and the evaluation value E takes a positive value, the weight vector W of the winner unit and the neighboring unit_iIs updated according to the above equation (5). In this case, the neighboring unit weight vector W_iIs not necessarily the winner unit weight vector W_iIt is not similar to. However, as learning progresses, the winner unit weight vector W_iAnd weight vector W of neighboring units_iGradually becomes similar.
[0049]
Although not particularly shown in the figure, when the evaluation with respect to the input learning vector I is negative, that is, when the evaluation value E takes a negative value, according to the above equation (6), the winner unit and the neighboring unit Weight vector W_iIs updated to move away from the learning vector I.
[0050]
FIG. 4A is a diagram illustrating a state of the weight vector space after learning progresses. As shown in the figure, when learning progresses to a considerable extent, the weight vector W_iEach point representing the end point of is gathered in the vicinity of the learning vector I given a positive evaluation. Therefore, this weight vector W_iBy examining the distribution of, the relationship between the first vector X and the second vector Y included in the learning vector I can be evaluated.
[0051]
For example, when an operation signal input to the control device is employed as the first vector X and an operation amount output from the control device is employed as the second vector Y, a data pair of these operation signal and operation amount and an individual evaluation Is sufficiently learned in advance by a neural network, as shown in FIG. 5A, a plurality of weight vectors W in a weight vector space can be obtained as a data pair of an operation signal and an operation amount capable of stably controlling a control amount._iWill be obtained as the end point of. For this reason, for example, the weight vector W shown in FIG._iBy interpolating the end points of these, a function representing the relationship between the operation signal and the operation amount as shown in FIG. As a result, by setting the function in the control device in advance, it is possible to easily design a control device that outputs an appropriate operation amount when given a certain operation signal. This also applies to determining the input / output characteristics of the target device.
[0052]
Further, as described above, the correlation between the first vector X and the second vector Y may be once evaluated after learning of the neural network, but it corresponds to the first vector X without performing such an evaluation. The preferred (which will give a positive evaluation) second vector Y is the weight vector W_iCan be calculated based on
[0053]
FIG. 5 shows the weight vector W_iFIG. 6 is a flowchart for explaining a procedure for calculating a preferable second vector Y corresponding to the first vector X from FIG.
[0054]
In this procedure, as shown in FIG. 6, only the first vector X is first given to the input layer 10 of the neural network. Here, the first vector X is distinguished from the first vector X which is a constituent element of the learning vector used in the learning stage.⁰(S201). In the following, this fact input vector X⁰In order to distinguish the second vector Y corresponding to the second vector Y which is a component of the learning vector used in the learning stage, in particular, the fact output vector Y⁰.
[0055]
Next, the fact input vector X⁰The output z of each neuron model on the competitive layer 12 for_p(Hereinafter referred to as “partial reactivity”) is calculated according to the following equation (7) (S202). In the equation (7), β is a given constant.
[0056]
[Formula 6]

Then, the partial reactivity z obtained in S202_pAnd combined load U_prAnd the fact output vector Y according to the following equation (8)⁰Is calculated (S203). FIG. 7 shows this fact output vector Y⁰This shows the concept of the calculation process.
[0057]
[Expression 7]

Equation (8) is the combined load U_prIs the partial reactivity z_pWhen the weighted addition is used, the fact output vector Y⁰The r-th component of can be calculated.
[0058]
For example, when the first vector X and the second vector Y both have only one component, if the weight vector W is represented on the weight vector space as shown in FIG. Vector X⁰, The weight vector W in the region 14 is obtained in the above-described processing._iCentered on the neuron model having the end point of_pWill have a large value. Then, when the equation (8) is calculated, the weight vector W in this region 14 is calculated._iThe value of the connection weight u for the neuron model having the end points of γ becomes dominant, and the fact output vector Y is shown around the arrow 16 in FIG.⁰Will be obtained.
[0059]
As described above, the weight vector W_iFrom this, a preferable second vector Y corresponding to the first vector X can be calculated.
[0060]
Embodiment 2. FIG.
In the second embodiment, an adaptive control apparatus to which the method according to the first embodiment described above is applied will be described.
[0061]
FIG. 9 is a diagram showing a configuration of the adaptive control apparatus according to the present embodiment. The adaptive control apparatus shown in the figure maintains and updates the neural network shown in FIG. 1 as hardware, and the first connection load storage unit 20 and the second connection are provided between the control unit 44 and the learning unit 46. The load storage unit 22 is shared and can be accessed from both sides at any time. The first connection load storage unit 20 stores the connection load w in the neural network shown in FIG._ijThe second connection weight storage unit stores the connection weight u in the neural network shown in FIG._ikIs memorized.
[0062]
First, the learning unit 46 includes a connection load update unit 28, an evaluation unit 38, and first and second connection load storage units 20 and 22. The evaluation unit 38 evaluates a control amount obtained from the control target 40 and gives an evaluation value E to the combined load update unit 28. Specifically, information representing the stability of the controlled variable can be adopted as the evaluation value E, and the value takes a continuous value from −1 to +1.
[0063]
An operation amount Y output from the control unit 44 is also input to the combined load update unit 28 via the switch 34 and the delay element 32. An operation signal X that is a difference between the reference input and the feedback amount is also generated by the adder 42 and is input to the coupling load update unit 28 via the delay element 30. The delay elements 30 and 32 are set with delay times so that the operation signal X, the operation amount Y, and the evaluation value E correspond to each other. The operation signal X has n data elements, the manipulated variable Y has m data elements, and these are transmitted through a plurality of signal lines.
[0064]
The combined load update unit 28 then calculates the weight vector W based on the above equations (5) and (6)._iIs updated in real time, and the updated combined load w_ij, U_ikAre stored in the first and second combined load storage units 20 and 22, respectively.
[0065]
On the other hand, the control unit 44 includes a partial reactivity calculation unit 24, an operation amount calculation unit 26, and first and second combined load storage units 20 and 22. The partial reactivity calculation unit 24 receives the bond load w from the first bond load storage unit 20._pqAnd the fact operation signal (actual input vector) X from the adder 42⁰Is entered. Then, in the partial reactivity calculation unit 24, based on these pieces of information, the partial reactivity z according to the above equation (7)._pAre calculated for neuron models on all competitive layers 12.
[0066]
Further, the operation amount calculation unit 26 receives the combined load u from the second combined load storage unit 22._prIs input from the partial reactivity calculation unit 24 to the partial reactivity z_pIs entered. Then, the manipulated variable calculator calculates the actual manipulated variable (actual output vector) Y according to the above equation (8) based on these pieces of information.⁰Is calculated. And the calculated actual manipulated variable Y⁰Is given to the control object 40.
[0067]
In the adaptive control apparatus described above, the contents of the first and second combined load storage units 20 and 22 are updated as needed based on various evaluation elements such as the stability of the control target 40. Then, according to the contents of the first and second combined load storage units 20 and 22 updated as needed, the fact operation signal X⁰The preferred manipulated variable Y (which will give a positive evaluation)⁰Is calculated by the control unit 44 and given to the control object 40. For this reason, even when the characteristics of the controlled object 40 change with time, it is possible to adaptively follow the change and maintain suitable control characteristics.
[0068]
The first and second embodiments of the present invention described above can be variously modified. For example, in the above description, the connection weight w of the neural network_ij, U_ikUpdate, ie, the weight vector W_iIs updated according to the equations (5) and (6), but other updating methods may be adopted.
[0069]
Evaluation value E_lAs E and E, any information corresponding to the learning vector can be adopted. For example, the confidence value, importance level, and stimulus level for each learning vector are set as the evaluation value E_lOr E.
[0070]
Furthermore, in the adaptive control device according to the second embodiment, a random value is given as a reference input to the adder 42, and the switch 34 is switched to the terminal 36 side to give a random value from the terminal 36. Prior to use, the neural network may be learned to some extent.
[Brief description of the drawings]
FIG. 1 is a diagram showing a configuration of a neural network used in a method according to a first embodiment of the present invention.
FIG. 2 is a flowchart for explaining a learning procedure of the neural network shown in FIG.
FIG. 3 is a diagram for explaining how the neural network shown in FIG. 1 learns.
FIG. 4 is a diagram showing a state of a weight vector space after learning.
FIG. 5 is a flowchart for explaining a procedure for calculating a fact output vector from a fact input vector using the neural network shown in FIG. 1;
6 is a diagram illustrating a state in which partial reactivity of each neuron model in a competitive layer with respect to a fact input vector is calculated using the neural network illustrated in FIG. 1;
FIG. 7 is a diagram showing how a fact output vector corresponding to a fact input vector is calculated based on partial reactivity of each neuron model in a competitive layer.
FIG. 8 is a diagram illustrating a process of calculating a fact output vector.
FIG. 9 is a diagram showing a configuration of an adaptive control apparatus according to Embodiment 2 of the present invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 10 Input layer, 12 Competitive layers, 20 1st bond load memory | storage part, 22 2nd bond load memory | storage part, 24 Partial reactivity calculation part, 26 Manipulation amount calculation part, 28 Bond load update part, 30, 32 Delay element, 34 Switch, 36 terminal, 38 evaluation part, 40 control object, 42 adder, 44 control part, 46 learning part.

Claims (7)

A device for determining a relationship between an operation signal and an operation amount in a control device,
Disposed in a first layer of the neural network, including an operation signal neuron model corresponding to the individual data elements that make up the operation signal, the neuron model for operation amount corresponding to the individual data elements that make up the operation amount, the A first layer neuron model group ;
A second layer neuron model group arranged in the second layer of the neural network and coupled to the motion signal and manipulated variable neuron models;
The first Soni Yuronmoderu group, enter the operation signal and the operation amount of the pair of learning sequence, including the evaluation value representing the degree of positive or negative evaluation to it a pair of said operating signal and the operation amount based on the individual evaluation information, a learning means for updating the connection weights between the second Soni Yuronmoderu group and said operation signal and the operation amount for the neuron model, of the second layer neuron model group For at least one neuron model, a weight vector having a connection load between the neuron model and each neuron model included in the first-layer neuron model group as an element, an operation signal for learning, and an operation amount. If the individual evaluation information includes a positive evaluation value for a learning vector that has individual data elements as constituent elements, the distance is proportional to the evaluation value. When the individual evaluation information includes a negative evaluation value, the calculation is performed to move away by a distance proportional to the evaluation value, and each element of the weight vector as a result of this calculation is the updated value of the associated load And learning means ,
A determining means based on the connection weights updated to determine the input and output characteristics of the control device, between each neuron model and the operation signal for the neuron model contained in the second layer neuron model group The input when the connection load is an input of the control device, and the connection load between each neuron model included in the second layer neuron model group and the neuron model for operation signals is the output of the control device, and the input A determination means for performing a calculation to obtain a function representing a relationship between the output and a function of a result of the calculation as an input / output characteristic of the control device;
Apparatus comprising: a. A control device using a neural network,
A neuron model for motion signals that is arranged in the first layer of the neural network and receives individual data elements constituting motion signals; and a neuron model for manipulated variables that receives individual data elements that constitute manipulated variables; A first layer neuron model group including:
A second layer neuron model group arranged in the second layer of the neural network and coupled to the motion signal and manipulated variable neuron models;
Based on the operation signal and the operation amount input to the first layer neuron model group, and individual evaluation information including an evaluation value representing a degree of positive or negative evaluation with respect to the operation signal and the operation amount, Learning means for updating a connection weight between a first layer neuron model group and the second layer neuron model group , wherein at least one neuron model in the second layer neuron model group is A weight vector having a connection load between each neuron model included in the first layer neuron model group as an element is a learning vector having individual data elements constituting the motion signal and an operation amount as elements. When the individual evaluation information includes a positive evaluation value, the individual evaluation information approaches a distance proportional to the evaluation value, and the individual evaluation information sets a negative evaluation value. No case performs an operation away by a distance that is proportional to the evaluation value, a value of the updated corresponding connection weights each element of the weight vector of the result of this calculation, the learning means,
Enter an operation signal to the operation signal for the neuron model, included in the operation signal and the bond and a preset value of the assignment to a result of the calculation in Equation load, the second layer neuron model group An operation amount calculation means for calculating a value based on the product of the connection load between a neuron model and the neuron model for operation as an operation amount to be output corresponding to the operation signal;
The control apparatus characterized by including. The control device according to claim 2,
The preset equation in the manipulated variable calculation means includes a vector having a connection load between one neuron model included in the second layer neuron model group and the neuron model for motion signal as an element , This is an expression that takes a larger value as the distance between the vector and the individual data elements constituting the signal is smaller,
The operation amount calculation means includes:
For each neuron model included in the second layer neuron model group, the connection load between the neuron model and the neuron model for motion signal, and the motion signal input to the neuron model for motion signal, Means for calculating a value of a result obtained by substituting into a preset equation as a partial reactivity with respect to the motion signal of each neuron model included in the second layer neuron model group ;
Based on the product of the connection weight between the neuron model included in the second layer neuron model group and the neuron model for manipulated variable and the partial reactivity of the neuron model included in the second layer neuron model group. Means for calculating the value as the manipulated variable ;
The control apparatus characterized by including. The control device according to claim 2 or 3,
The individual evaluation information is generated by evaluating a control amount acquired from the control target when the operation amount is given to the control target of the control device. A data generation device that generates second data based on first data,
A first neuron model disposed in a first layer of the neural network and corresponding to individual data elements of the first data; and a second neuron model corresponding to individual data elements of the second data. Layer neuron model group,
A second layer neuron model group disposed in the second layer of the neural network and coupled to the first and second neuron models;
First data is input to the first neuron model, and a connection load between the first layer neuron model group and the second layer neuron model group and the first data are substituted into a preset equation. Data calculating means for calculating, as second data , a value based on the product of the result of the operation and the connection load between the second layer neuron model group and the second neuron model ;
In the learning stage, the first and second data pairs for learning are sequentially input to the first layer neuron model group, and the first and second data pairs for learning and positive or negative evaluations thereof are performed. The connection load between the first layer neuron model group and the second layer neuron model group is updated based on the individual evaluation information including the evaluation value indicating the degree of the second layer neuron model . In this update process, the second layer For at least one neuron model in the neuron model group, a weight vector having a connection weight between the neuron model and each neuron model included in the first layer neuron model group as an element is used for the learning. When the individual evaluation information includes a positive evaluation value with respect to a learning vector having individual data elements of the first and second data as elements, the evaluation value is proportional to the evaluation value. When the individual evaluation information includes a negative evaluation value, the calculation is performed to move away by a distance proportional to the evaluation value, and after updating the corresponding combined load, each element of the weight vector as a result of this calculation is performed. The value of
A data generation apparatus characterized by that. An input / output characteristic determination device for determining a relationship between input data and output data of a target device,
Disposed in a first layer of the neural network, the first containing the input data neuron model corresponding to the individual data elements of the input data, and the neuron model output data corresponding to the individual data elements of said output data, the Layer neuron model group ,
A second layer neuron model group arranged in the second layer of the neural network and coupled to the input data and output data neuron models ;
Before SL sequentially inputs a pair of input and output data for learning first Soni Yuronmoderu group, to the individual evaluation information including an evaluation value representing the degree of positive or negative evaluation to it and said pair of data based on, a learning means for updating the connection weights between the input data and output data for a neuron model second Soni Yuronmoderu group, at least one neuron of the second layer neuron model group For the model, a weight vector having each element as a connection weight between the neuron model and each neuron model included in the first layer neuron model group, and each data element of the input data and output data as each element When the individual evaluation information includes a positive evaluation value with respect to the learning vector to be When the valence information includes a negative evaluation value, a calculation is performed to move away by a distance proportional to the evaluation value, and each element of the weight vector as a result of this calculation is used as a value after updating the corresponding connection weight. Means ,
A determining means for determining the input-output characteristics of the target device based on the previous SL connection weights, the connection weight between the neurons model and the input data for the neuron model contained in the second layer neuron model group The input and the output when the connection load between each neuron model included in the second layer neuron model group and the neuron model for output data is the output of the target device. A determination means for performing a calculation to obtain a function representing a relationship between the two, and setting a function of a result of the calculation as an input / output characteristic of the target device;
Output characteristic determining apparatus comprising: a. A correlation evaluation apparatus for evaluating a correlation between first data and second data,
A first data neuron model corresponding to each data element of the first data, and a second data neuron model corresponding to each data element of the second data , arranged in the first layer of the neural network ; A first layer neuron model group including :
Wherein arranged in the second layer of the neural network, a second layer neuron model group that binds to said first data and second data neuron model,
Before SL sequentially inputs a pair of first and second data for learning the first Soni Yuronmoderu group, individual evaluation information including an evaluation value representing the degree of positive or negative evaluation to it and said pair of data based on the bets, a learning means for updating the connection weights between the second Soni Yuronmoderu group and the first data and the second data neuron model, of said second layer neuron model group For at least one neuron model, a weight vector having each element as a connection weight between the neuron model and each neuron model included in the first layer neuron model group is used as the first and second data for learning. When the individual evaluation information includes a positive evaluation value with respect to a learning vector having individual data elements as elements, the distance is proportional to the evaluation value, When the evaluation information includes a negative evaluation value, a calculation is performed to move away by a distance proportional to the evaluation value, and each element of the weight vector as a result of the calculation is used as a value after updating the corresponding joint weight. Means ,
Correlation an evaluation means for evaluating the respective neuron model and the first data for neuron models included in the second layer neuron model group between the first and second data based on the previous SL coupling load Is the first data, and the connection load between each neuron model included in the second layer neuron model group and the second data neuron model is the second data. An evaluation means for performing an operation for obtaining a function representing a relationship between the data and the second data, and using a function as a result of the operation as a function representing a correlation between the first data and the second data; ,
Correlation evaluation device, characterized in that it comprises a.

Images