EP1588229A2 - Verfahren zur ermittlung des zulässigen arbeitsbereichs eines neuronalen netzes - Google Patents
Verfahren zur ermittlung des zulässigen arbeitsbereichs eines neuronalen netzesInfo
- Publication number
- EP1588229A2 EP1588229A2 EP04700118A EP04700118A EP1588229A2 EP 1588229 A2 EP1588229 A2 EP 1588229A2 EP 04700118 A EP04700118 A EP 04700118A EP 04700118 A EP04700118 A EP 04700118A EP 1588229 A2 EP1588229 A2 EP 1588229A2
- Authority
- EP
- European Patent Office
- Prior art keywords
- input data
- simplex
- neural network
- point
- data record
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
Definitions
- the invention relates to a method for checking whether an input data set is within the permissible working range of a neural network, and to a corresponding computer program product and system.
- Neural networks are used for data-driven modeling, for example for physical, biological, chemical and technical processes and systems, cf. Babel.: Possible uses of neural networks in industry: Pattern recognition using supervised learning processes - with examples from traffic and medical technology, Expert Verlag, Renningen-Malmsheim, 1997. In particular, the areas of application of neural networks include process optimization, image processing, pattern recognition, robot control and medical technology.
- a neural network Before a neural network can be used for forecasting or optimization purposes, it must be trained.
- the weights of the neurons are usually adjusted using an iterative process based on training data, cf. Betzmann F .: Process modeling: Modeling of continuous systems with neural networks, website NN-Tool, www.baermann.de and Bärmann F .: Neural networks. Lecture notes. FH-Gelsenmaschinen, Department of Physical Technology, Department of Neuroinformatics, 1998.
- DE 195 31 967 discloses a method for training a neural network with the non-deterministic behavior of a technical system.
- the neural network is integrated into a control loop in such a way that the neural network outputs a manipulated variable to the technical system as an output variable and the technical system generates a controlled variable from the manipulated variable supplied by the neural network, which is supplied to the neural network as an input variable.
- the manipulated variable is overlaid with a noise of known noise distribution before it is fed to the technical system.
- Further methods for training neural networks are known from DE 692 28 412 T2 and DE 198 38 654 Cl.
- EP 0 762 245 B1 also describes a method for the detection of faulty ones Predictions known in a neuromodel-based or neural process control.
- a common disadvantage of these methods known from the prior art is that they can only provide information about the sensitivity of the model made available by the neural network with regard to variations in the training data. However, it is not possible to make a statement about the trustworthiness of a forecast created by the neural network.
- the invention is therefore based on the object of providing a method which makes it possible to check whether an input data record lies in the permissible working range of a neural network. Furthermore, the invention is based on the object of creating a corresponding computer program product.
- the present invention makes it possible to check an input data record for a neural network to determine whether it is within the permissible working range of the neural network
- the invention is based on the knowledge that no structural information flows into neural networks, but rather only training input data sets are used which have been determined, for example, by measurement technology. Because of this, such models can only provide reliable forecasts in the areas in which the models have been trained.
- Hybrid models can be extrapolated as an overall model, but for each individual data driven sub-component, that is, for the neural networks included in the hybrid model, the interpolation range must be checked.
- the working area of a neural network is defined by the convex envelope spanned by the training input data records of the neural network.
- a neural network has a number of inputs and a number of b outputs.
- data records for the a input and the b output parameters are metrologically recorded.
- the input parameters can be data relating to the raw materials used, their composition and / or parameters of the production system, such as pressures, temperatures and the like.
- the resulting product properties are then measured for the output parameters, for example. In this way you get training data sets, each one
- the following definition of the convex hull is used to determine the permissible working range:
- P be a given, finite set of n points p ⁇ t ... j? n .
- a point x i.e. a certain one
- the immediate vicinity of the convex envelope is also still considered to be a permissible work area, since neural networks can still provide meaningful results even in the immediate vicinity of the convex envelope.
- the working area is limited to the convex hull, since an exact statement about where the "immediate neighborhood" ends cannot be made. In particular for critical applications, for example relating to continuous production, the working area is therefore on the inside of the convex hull, with the external environment in the immediate vicinity of the convex hull from
- the invention can be resorted to three fundamentally different, very efficient methods.
- a simplex is first formed from a number of d + 1 non-collinear points from the set P, where d is the prostitute sion of the space spanned by P.
- a point is then selected from inside this simplex.
- the center of gravity of the simplex can be used, which is calculated from the corner points of the simplex. This point is referred to below as x 0 .
- the distance [x, x 0 ] between the point x defined by the input data set and the point x selected from the simplex is considered. Then it is checked whether there is an intersection of the segment [x, x 0 ] with a facet of the simplex.
- the facets are the "side surfaces" of the Simplex.
- the test as to whether the formation of a further simplex, which includes a section of the segment [x, x 0 ], is carried out as follows: First the corner points of the facet, the is cut from the distance [x, x 0 ]. Then another point is selected from the set P. This can be any point that does not belong to the corner points of the facet.
- this experimentally formed further simplex contains a section of the segment [x, x 0 ], then this experimentally formed further simplex is used as a simplex for a further iteration of the method.
- the further point selected from P is replaced by another point in order to form a further simplex on a trial basis and in order to check again whether a section of segment [x, x 0 ] lies in the trial-formed simplex perform.
- Simplex is possible that includes a section of the segment [x, x 0 ], ie x lies outside the convex hull.
- a different geometric property of the convex envelope is used. This property reads:
- a module for checking whether an input data record is in a permissible working range of the neural network is connected upstream of a neural network. If the system in question is a system with several neural networks and / or a system with rigorous model components, that is to say a so-called hybrid model, such a module is preferably connected upstream of each neural network of the system. If several neural networks are used, these modules can be linked with a logical "AND" in order to ensure that an input data record is within the permissible working range of all of these neural networks. This is particularly important in hybrid models.
- FIG. 1 shows a flow diagram of a first embodiment of a method for checking whether an input data record lies in the convex envelope
- FIG. 2 shows a development of the method in FIG. 1 for determining a further simplex
- FIG. 3 shows a further embodiment of a method according to the invention for checking whether an input data record lies in the convex envelope
- FIG. 4 shows a graphical illustration of the method in FIG. 3
- Figure 5 shows another embodiment of the method for checking whether a
- Input data record lies in the convex hull, based on a check whether there is a solution for the system of equations given by the analytical definition of the convex hull,
- Figure 6 is a block diagram of an embodiment of an inventive
- FIG. 1 illustrates a first embodiment of the method for checking whether an input data record lies in the convex envelope. This method starts from a point x 0 inside the convex hull and checks whether the distance [x, x 0 ] lies inside the convex hull.
- x is the point determined by the input sentence and you want to know whether it is also inside the convex hull.
- ⁇ ) : ⁇ ⁇ + c ⁇ ( .
- x,: a convex linear combination for a point x t € conv (P) that is closer to x than x 0 . If one assumes the linear combination described above for xo, in which the highest d + 1 coefficients are not equal to 0, and one chooses the largest possible c, one obtains in this way the intersection of the segment [xo, *] with a facet of the simplex, which is spanned by the points from P belonging to the coefficients.
- the algorithm delivers a convex linear combination to represent the point. If the point lies outside, d points are obtained by which a hyperplane E is determined, which separates the point set P from the point x. This means that all points of R d that lie on the same side of E as point x do not belong to the convex hull. can. This can be used in a multiple evaluation to significantly speed up the entire evaluation.
- FIG. 1 One form of implementation of this method is illustrated in FIG. 1:
- step 100 an input data record for which a forecast is to be created is entered.
- This input data record for the neural network determines a point x.
- step 101 a number of d + 1 non-collinear points from the
- step 102 the index / is set to zero.
- step 103 a simplex S 1 is formed from the points selected in step 101.
- step 104 a point x, from the inside of the simplex S ; selected.
- the center of gravity is calculated from the corner points of the Simplex S in order to obtain the point x t .
- step 105 a distance [x / x] between x and x is defined.
- step 106 it is checked whether an intersection x / + 1 of the distance [x / x] with a facet of the simplex S s lies between x and x. It is therefore checked whether, starting from x, starting on the straight line direction x, first x or a facet of the Simplex S is reached.
- step 108 checks whether it is possible to find a further simplex S M in P which contains both the intersection x / + 1 and a section of the straight line g. If this is not possible, it is output in step 109 that x lies outside the convex hull.
- Step 106 performed again with respect to the further simplex.
- FIG. 2 shows a further development of the method of FIG. 1 for carrying out the test in step 108.
- the corner points of the facet of S on which the intersection point x / + 1 lies are first determined in step 200.
- a further point is selected from P which is not already a corner point of the facet of S, and which is not collinear with the corner points of the facet.
- step 202 a simplex S 'is formed from the corner points and the further point from P.
- step 203 it is checked whether the simplex S 'contains a section of the route [x / x]. If this is the case, in step 204 the further simplex S M sought is set equal to the simplex S '. This then also answers the question of whether it is actually possible to form such a Simplex S / + 1 . If the check in step 203 shows that the simplex S 'contains no section of the straight line g, it is checked in step 205 whether all the points in question from P have already been selected in step 201. If this is not the case, a further point from P is selected in step 201, which has not yet been selected in order to carry out a further iteration of the method.
- step 206 If no simplex S ; +1 could be found even after “trying out” all the points in question from P, appropriate information is output in step 206. This also means that point x lies outside the convex hull.
- a particular advantage of this embodiment is that the method always leads to a statement after a finite number of steps as to whether the input data record is in the convex envelope and thus in the work area or not.
- FIG. 3 shows a further embodiment of a method for checking whether an input data record lies in the convex envelope. This procedure does not result directly from the definition of the convex hull as a linear combination of the support points. Rather, another geometric property of the convex hull is used here, which is also illustrated graphically in FIG. 4:
- Ti pi - x are the location vectors of the data points in a coordinate system that originates from the data point to be examined.
- the inequality can be queried to "greater”, since the normal vector -k represents the same hyperplane as k. Points on the facets of the convex envelope lead to a scalar product equal to 0 and are therefore part of the convex envelope.
- An optimization method is preferably used for the search for a hyperplane.
- the point to be examined lies outside the convex hull. No hyperplane for which ⁇ 0 applies can be found for points within the convex hull.
- Various methods can be used as an optimization method, such as the MATLAB routine fminsearch, as well as the gradient method, a Levenberg-Marquard algorithm or an evolution strategy, which can also be used in combination with local methods.
- a major advantage for the runtime behavior of the algorithm is that if a corresponding hyperplane has been found for a data point, this too - 17 -
- FIG. 3 illustrates this method using a flow chart.
- step 300 the input data record, that is the point, is entered.
- Step 301 uses one or more of the methods mentioned to check whether there is a hyperplane which contains x and for which k - ⁇ > 0, i - l, ..., n applies, where k is the normal vector of the hyperplane being sought and r i is the difference vector between a point p t and x given by a training input data record.
- step 302 If there is such a hyperplane, it follows in step 302 that x lies in the convex hull. In the opposite case, information is output in step 303 that x lies outside the convex hull.
- the check in step 301 as to whether there is a suitable hyperplane is illustrated in FIG. 4.
- the points p t located in the shaded area in FIG. 4 span a convex envelope 400.
- the point x is outside the convex hull 400.
- Between the point x and the points p t there are the difference vectors r t p l , - x.
- a hyperplane 401 runs through x and is described by the normal vector k. Since all points p. the convex hull 400 on the same side of the
- FIG. 5 illustrates a further method for checking whether an input data record x lies in the convex envelope.
- Equation 3 we transform Equation 3 by multiplying it by a matrix M on both sides.
- x: M -x.
- R (,) is the transpose of the matrix R w . If all components of the coefficient vector ⁇ found in this way meet the constraints ⁇ , ⁇ 0, a convex linear combination has been found for the point x and the point x is therefore inside the convex hull. Otherwise, we set all coefficients that violate the constraint to zero for the rest of the procedure and try to correct the components that do not violate the constraint so that this step is compensated for. In practice, this is accomplished by eliminating all components that violate the constraint from the vector ⁇ and all the associated columns from the matrix P. We denote the smaller dimension vector obtained in this way and the matrix with fewer columns obtained with ⁇ (, + 1) and R (, + 1) .
- the method comes to a result after a maximum of n steps.
- FIG. 5 An embodiment of this method is illustrated in FIG. 5.
- step 500 the index i is set to zero.
- step 501 a starting value for the n-dimensional vector ⁇ ⁇ 0 'that fulfills the constraints is selected.
- ⁇ t 11 n can be selected, for example.
- step 502 the matrix M is calculated. Based on this, the matrix P ⁇ and the vectors and x ( ' ⁇ are calculated in step 503.
- ⁇ ⁇ (0 + P ii) T ⁇ (x - x () ) is calculated in step 504.
- step 508 the index is incremented to perform another iteration of the method.
- FIG. 6 shows a block diagram of an embodiment of a system 600 according to the invention.
- the input module 601 is linked to a module 602, which is used to check whether an input data record lies within the convex envelope of the neural network 603. This check is carried out, for example, according to one of the related to the
- FIGS 1 to 5 described method or by another method.
- the module 602 is linked to the neural network 603. If module 602 determines that an input data record lies within the permissible working range of the neural network, which is given by the convex envelope, this is input
- system 600 can also include further neural networks (hybrid model), each of which is preceded by a module corresponding to the module 602.
- hybrid model the results of the individual modules 602 must then be linked with a logical “AND”. This ensures that all neural networks of the hybrid model 600 for a specific input data record of the input module 601 can be operated in a permissible working range.
- system 600 can also contain rigorous model components.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Complex Calculations (AREA)
- Computer And Data Communications (AREA)
- Input From Keyboards Or The Like (AREA)
Abstract
Description
Claims
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| DE10301420A DE10301420A1 (de) | 2003-01-16 | 2003-01-16 | Verfahren zur Ermittlung des zulässigen Arbeitsbereichs eines neuronalen Netzes |
| DE10301420 | 2003-01-16 | ||
| PCT/EP2004/000019 WO2004063832A2 (de) | 2003-01-16 | 2004-01-05 | Verfahren zur ermittlung des zulässigen arbeitsbereichs eines neuronalen netzes |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| EP1588229A2 true EP1588229A2 (de) | 2005-10-26 |
| EP1588229A3 EP1588229A3 (de) | 2005-11-02 |
Family
ID=32602600
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP04700118A Withdrawn EP1588229A3 (de) | 2003-01-16 | 2004-01-05 | Verfahren zur ermittlung des zulässigen arbeitsbereichs eines neuronalen netzes |
Country Status (4)
| Country | Link |
|---|---|
| US (1) | US20040172375A1 (de) |
| EP (1) | EP1588229A3 (de) |
| DE (1) | DE10301420A1 (de) |
| WO (1) | WO2004063832A2 (de) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20050251376A1 (en) * | 2004-05-10 | 2005-11-10 | Alexander Pekker | Simulating operation of an electronic circuit |
| EP1793243A1 (de) * | 2005-12-05 | 2007-06-06 | Leica Geosystems AG | Verfahren zur Auflösung einer Phasenmehrdeutigkeit |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5835901A (en) * | 1994-01-25 | 1998-11-10 | Martin Marietta Corporation | Perceptive system including a neural network |
| US5835907A (en) * | 1995-12-20 | 1998-11-10 | Mci Communications Corporation | Emergency PCS system for identification and notification of a subscriber's location |
| US7203716B2 (en) * | 2002-11-25 | 2007-04-10 | Simmonds Precision Products, Inc. | Method and apparatus for fast interpolation of multi-dimensional functions with non-rectangular data sets |
-
2003
- 2003-01-16 DE DE10301420A patent/DE10301420A1/de not_active Withdrawn
-
2004
- 2004-01-05 EP EP04700118A patent/EP1588229A3/de not_active Withdrawn
- 2004-01-05 WO PCT/EP2004/000019 patent/WO2004063832A2/de not_active Ceased
- 2004-01-15 US US10/758,322 patent/US20040172375A1/en not_active Abandoned
Non-Patent Citations (1)
| Title |
|---|
| See references of WO2004063832A2 * |
Also Published As
| Publication number | Publication date |
|---|---|
| US20040172375A1 (en) | 2004-09-02 |
| WO2004063832A3 (de) | 2005-09-09 |
| WO2004063832A2 (de) | 2004-07-29 |
| EP1588229A3 (de) | 2005-11-02 |
| DE10301420A1 (de) | 2004-07-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| DE69324052T2 (de) | Neuronalnetzwerk-Lernsystem | |
| EP2112568B1 (de) | Verfahren zur rechnergestützten Steuerung und/oder Regelung eines technischen Systems | |
| EP3785177B1 (de) | Verfahren und vorrichtung zum ermitteln einer netzkonfiguration eines neuronalen netzes | |
| DE112016001796T5 (de) | Feinkörnige bildklassifizierung durch erforschen von etiketten von einem bipartiten graphen | |
| DE102018111905A1 (de) | Domänenspezifische Sprache zur Erzeugung rekurrenter neuronaler Netzarchitekturen | |
| DE69224718T2 (de) | Klassifizierungsverfahren für Rechnerarchitekturen | |
| DE112021000251T5 (de) | Verfahren zum auswählen von datensätzen zum aktualisieren eines moduls mit künstlicher intelligenz | |
| DE69802372T2 (de) | Klassifizierungssystem und -verfahren mit N-Tuple- oder RAM-basiertem neuronalem Netzwerk | |
| DE69328596T2 (de) | Optimierung eines Neuralnetzwerks mit Vorwärtskopplung | |
| DE10023377C2 (de) | Verfahren zur Erhöhung der Leistungsfähigkeit einer Computereinrichtung bei Finite-Elemente-Simulationen und eine solche Computereinrichtung | |
| DE19703964C1 (de) | Verfahren zur Transformation einer zur Nachbildung eines technischen Prozesses dienenden Fuzzy-Logik in ein neuronales Netz | |
| EP1588229A2 (de) | Verfahren zur ermittlung des zulässigen arbeitsbereichs eines neuronalen netzes | |
| DE102021202335A1 (de) | Verfahren und Vorrichtung zum Prüfen eines technischen Systems | |
| DE102024201461A1 (de) | Verfahren zum Optimieren eines herzustellenden Gyroskop-Systems | |
| WO1998007100A1 (de) | Rechnergestütztes verfahren zur auswahl von trainingsdaten für ein neuronales netz | |
| DE102020213238A1 (de) | Erzeugung von vereinfachten computer-implementierten neuronalen netzwerken | |
| DE19831651C1 (de) | Verfahren zum Erzeugen eines regel- und anpassbaren Netzwerkes von Modellen von Verhaltensmustern einschließlich Software-Systemen | |
| DE102020212328A1 (de) | Vorrichtung und computerimplementiertes Verfahren für eine Netzwerkarchitektursuche | |
| DE102018109691A1 (de) | Verfahren zur computerunterstützten Fertigungsoptimierung zumindest eines Fertigungsschritts | |
| DE102023211078A1 (de) | Verfahren sowie ein System zum Optimieren einer Netzwerkarchitektur eines künstlichen neuronalen Netzwerks | |
| DE102022201718A1 (de) | Computer-implementiertes Verfahren zum Erstellen eines Gültigkeitsmodells zum Ermitteln einer Modellierungsunzuverlässigkeit in einem Steuergerät mit begrenzter Rechenkapazität mithilfe maschineller Lernverfahren | |
| DE102023206789A1 (de) | Verfahren zum Training eines Faltungs-Neuronalnetzes, das in Schichten angeordnete Knoten und eine Beschneidungsmaske umfasst | |
| DE102025102861A1 (de) | Eine numerische Methode zur Lösung der Schrödinger-Gleichung in hochdimensionalen Räumen | |
| EP1190383B1 (de) | Verfahren zur rechnergestützten ermittlung einer zugehörigkeit einer vorgegebenen eingangsgrösse zu einem cluster | |
| DE112023005296T5 (de) | Fähigkeitenerwerb für dynamische behandlungskonzepte weitere anwendungsinformationen |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
| PUAK | Availability of information related to the publication of the international search report |
Free format text: ORIGINAL CODE: 0009015 |
|
| AK | Designated contracting states |
Kind code of ref document: A2 Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IT LI LU MC NL PT RO SE SI SK TR |
|
| AX | Request for extension of the european patent |
Extension state: AL LT LV MK |
|
| AK | Designated contracting states |
Kind code of ref document: A3 Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IT LI LU MC NL PT RO SE SI SK TR |
|
| AX | Request for extension of the european patent |
Extension state: AL LT LV MK |
|
| RIC1 | Information provided on ipc code assigned before grant |
Ipc: 7G 06N 3/08 A |
|
| 17P | Request for examination filed |
Effective date: 20060309 |
|
| DAX | Request for extension of the european patent (deleted) | ||
| RBV | Designated contracting states (corrected) |
Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IT LI LU MC NL PT RO SE SI SK TR |
|
| RIN1 | Information on inventor provided before grant (corrected) |
Inventor name: HUEBL, PETER Inventor name: MRZIGLOD, THOMAS Inventor name: MOGK, GEORG |
|
| 17Q | First examination report despatched |
Effective date: 20110803 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN |
|
| 18D | Application deemed to be withdrawn |
Effective date: 20120214 |