CN103745079A - Curve fitting method based on abstract convex estimations - Google Patents

Curve fitting method based on abstract convex estimations Download PDF

Info

Publication number
CN103745079A
CN103745079A CN201310600939.7A CN201310600939A CN103745079A CN 103745079 A CN103745079 A CN 103745079A CN 201310600939 A CN201310600939 A CN 201310600939A CN 103745079 A CN103745079 A CN 103745079A
Authority
CN
China
Prior art keywords
formula
curve
lower bound
leaf node
matrix
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.)
Pending
Application number
CN201310600939.7A
Other languages
Chinese (zh)
Inventor
张贵军
周晓根
郝小虎
张贝金
明洁
刘玉栋
秦传庆
陈铭
夏华栋
李栋炜
梅珊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN201310600939.7A priority Critical patent/CN103745079A/en
Publication of CN103745079A publication Critical patent/CN103745079A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Complex Calculations (AREA)

Abstract

A curve fitting method based on abstract convex estimations includes the following steps that supporting vectors begin to be built from two boundary points to form an initial supporting matrix, and the initial supporting matrix is used as a root to built a binary tree; supporting vectors are built for known data points, and the binary tress is updated according to a certain condition; a sampling step length is set, sampling is conducted within a definitional domain range of x, tree leaf child nodes to which sampling points belong are found out, and lower bound low estimators of a tree leaf child node zone to which the sampling points belong are calculated; all the sampling points are connected, and then lower bound saw tooth estimations of a curve to be fitted can be obtained; upper bound saw tooth estimations of the curve to be fitted can be obtained through the same method, and the curve which is obtained by averaging the upper bound saw tooth estimations and the lower bound saw tooth estimations is used as the curve to be fitted. The curve fitting method based on the abstract convex estimations has the advantages that calculated amount is small, the original data points can be reserved, and fitting efficiency and the accuracy rate are high.

Description

A kind of curve-fitting method of estimating based on abstract convex
Technical field
The present invention relates to a kind of mathematical modeling, computer application field, in particular, a kind of curve-fitting method of estimating based on abstract convex.
Background technology
Science and engineering problem can obtain some discrete data by methods such as sampling, experiment, and according to these data, we often wish to obtain a continuous curve and given data matches, and this process is just called curve.Curve application is very extensive, occupies very consequence in computational science.For example, the field of people to a certain the unknown, in order to explore its inherent rule, holds its developing direction, will use curve.Therefore, comprehensive research of curve-fitting method is had to positive realistic meaning.At present, curve fitting technique has all obtained wide application in the fields such as processes and displays of image processing, reverse-engineering, computer-aided design (CAD) and test data.The method of curve has a lot, and conventional method has least square method and method of interpolation etc.
The least square method of curve is according to known data (x k, y k) (k=1,2 ..., K) (x wherein kfor measured quantity, y kfor its corresponding functional value, the quantity size size that K is data point) choose an approximate function f (x), make E 2(f) (root-mean-square error) minimum; Least square fitting great advantage is that fitting precision can regulate and control, can select different fitting precisions according to actual service condition, but in actual applications, some occasion need to retain former data point, at this moment use the common methods such as least square method just improper.The method of interpolation of curve is according to interpolation principle structure polynomial of degree n P n(x), make P n(x) in the data of each test point, just in time pass through eyeball; Although method of interpolation can retain former data point,, in the ordinary course of things, we have gathered a lot of sampling points in order to reflect actual conditions as far as possible, have caused interpolation polynomial P n(x) number of times is very high, and this has not only increased calculated amount, and has affected fitting precision.
Therefore, existing technology exists defect aspect curve, needs to improve.
Summary of the invention
In order to overcome existing method, can not retain former data point, calculated amount is large, fitting precision is lower deficiency, the invention provides a kind of when guaranteeing fitting precision, can retain former data point, calculated amount is little, fitting precision the is higher curve-fitting method based on abstract convex estimation.
The technical solution adopted for the present invention to solve the technical problems is:
A curve-fitting method of estimating based on abstract convex, described curve-fitting method comprises the following steps:
1) parameter initialization: constant M is set, input data (x k, y k), k=1,2 ..., K, the quantity size size that wherein K is data point;
2) boundary condition initialization, process is as follows;
2.1) according to existing data point, determine the field of definition [a, b] of variable x, wherein a is the lower bound of variable x, and b is the upper bound;
2.2) model conversion: former variable x is done to following linear transformation:
Figure BDA0000420595640000021
X wherein 1', x 2' be linear transformation variable;
Formula (1) is converted, obtains:
x=x 1′(b-a)+a (2)
2.2.1) establish
Figure BDA0000420595640000022
wherein,
Figure BDA0000420595640000023
the linear transformation variable 1 that represents the 1st data,
Figure BDA0000420595640000024
the linear transformation variable 1 of the 2nd data;
2.2.2) right according to formula (2)
Figure BDA0000420595640000025
be converted to former variable x 1, x 2, x wherein 1=b, x 2=a;
2.3) calculate initial support vector sum proppant matrix, and set up tree:
2.3.1) according to formula (4) respectively calculation level (x ' 1, y b) and (x ' 2, y a) support vector,
l k = ( x 1 k f ( x k ) , x 2 k f ( x k ) , . . . , x N + 1 k f ( x k ) ) - - - ( 4 )
Wherein, y brepresent functional value corresponding to upper bound b, y arepresent functional value corresponding to lower bound a, the x in formula kwith x ' replacement, f () replaces with y;
2.3.2) two support vectors of gained are formed to proppant matrix L, as shown in Equation (7):
Figure BDA0000420595640000027
2.3.3) take proppant matrix L as root foundation tree;
3) calculate the support vector of each data point and upgrade tree, each data point is handled as follows:
3.1) according to formula (1) to x kas linear transformation, be
3.2) according to formula (4) calculate point (x ' k, y k) support vector l k, the x in formula kwith x ' kreplace, f () uses y kreplace;
3.3) according to conditional relationship formula (5) and (6), upgrade tree:
∀ i , j ∈ I , i ≠ j : l i k i > l i k j - - - ( 5 )
∀ v ∈ Λ k \ L , ∃ i ∈ I : l i k i ≤ v i - - - ( 6 )
V ∈ Λ wherein kl represent that v belongs to Λ kbut do not belong to L,
Figure BDA0000420595640000038
represent to exist;
3.3.1) find for step 3.2) the support vector l that produces kthe leaf node of (6), i.e. v do not satisfy condition i=l k;
3.3.2) use l kreplacement step 3.3.1) i support vector in the leaf node matrix finding in
Figure BDA0000420595640000034
thereby form new leaf node;
3.3.3) the new leaf node producing the determining step 3.3.2) relational expression (5) that whether satisfies condition, if met, retains, otherwise deletes;
4) calculate Lower Bound Estimation value;
4.1) sampling step length is set, and sampling in field of definition [a, b] scope;
4.2) to step 4.1) each sampled point x cbe handled as follows:
4.2.1) according to formula (1), do linear transformation, obtain new vector x c'=(x c1', x c2');
4.2.2) according to formula (10) from step 3.3) tree find and comprise vector x c' leaf node, in formula
Figure BDA0000420595640000035
use x c' replace, if leaf node meets formula (10), comprise, otherwise do not comprise;
x j k j x ^ i r > x i k j x ^ j r , i , j ∈ I , i ≠ j - - - ( 10 )
Wherein,
Figure BDA0000420595640000037
for the element in looked for leaf node matrix;
4.2.3) according to formula (3), obtain vector x cthe Lower Bound Estimation value y of the leaf node corresponding region at ' place l';
4.2.5) due in step 2.3) in added a constant M while being converted into IPH, in the hope of Lower Bound Estimation value should deduct M, i.e. y l=y l'-M;
4.2.6) output Lower Bound Estimation value y land corresponding sampled point x c;
5) calculate Estimation of Upper-Bound value, process is as follows:
5.1) the y in data point kget negative;
5.2) by step 2) to 4) calculating Lower Bound Estimation value;
5.3) to step 5.2) the Lower Bound Estimation value of gained gets negatively, is Estimation of Upper-Bound value y u;
5.4) output Estimation of Upper-Bound value y uand corresponding sampled point x c;
6) curve, process is as follows:
6.1) connect each point (x c, y l) obtain treating the Lower Bound Estimation of matched curve;
6.2) connect each point (x c, y u) obtain treating the Estimation of Upper-Bound of matched curve;
6.3) Estimation of Upper-Bound and Lower Bound Estimation being averaged to the curve obtaining is and treats matched curve.
Further, the constant M in step 1), if function f meets following two conditions:
i ) ∀ x , y ∈ R + n , x ≥ y ⇒ f ( x ) ≥ f ( y )
ii ) ∀ x ∈ R + n , ∀ λ ∈ R + + : f ( λx ) = λf ( x )
Wherein,
Figure BDA00004205956400000410
represent arbitrarily,
Figure BDA0000420595640000043
Figure BDA0000420595640000044
claim function f for just homogeneous increasing function IPH;
Add one enough large constant M can will treat that function corresponding to matched curve is converted to IPH function, adds M to each y in data.
Further, the method for calculating support vector and proppant matrix step 2.3) is:
Suppose that the field of definition of N dimension objective function utilized model conversion formula (1) unit's of being mapped to simplex space
Figure BDA0000420595640000045
suppose that it has K support function, the point (x based on given k, f (x k)) (k=1,2 ..., K), the Lower Bound Estimation of objective function f is provided by following formula:
H K ( x ) = max k ≤ K min i = 1 , . . . N + 1 x i l i k - - - ( 3 )
Wherein max represents maximum, and min represents minimum, x ifor the vector in unit simplex space;
K support function is:
h k ( x ) = min i = 1 , . . . N + 1 x i l i k = f ( x k ) min { x 1 x 1 k , . . . , x N + 1 x N + 1 k }
Wherein, l kbe called support vector, as the formula (4).Consider a set that contains K support vector
Figure BDA0000420595640000048
make I={1,2 ..., N+1}, N+1 the support vector of satisfy condition (5) and (6)
Figure BDA0000420595640000049
be called an effective proppant matrix;
Suppose an effective proppant matrix as the formula (7):
According to this proppant matrix, can calculate corresponding lower bound support function extreme value x minthe Lower Bound Estimation value d corresponding with it:
x min(L)=diag(L)/Trace(L) (8)
d(L)=H k(x min)=1/Trace(L) -1 (9)
Wherein, diag is the positive diagonal entry of matrix, and Trace is matrix trace, i.e. the positive diagonal entry sum of matrix.
Further again, step 4.2.2) in, find out and comprise vector x c' the method for leaf node be:
Suppose
Figure BDA0000420595640000052
for H k(x) r local minimum, the relational expression that satisfies condition (5) of its correspondence and the proppant matrix of (6) are L r, by formula (9), had lower Bound Estimation value
H k ( x ^ r ) = 1 / Trace ( L r ) - 1
Figure BDA0000420595640000055
i element be
x ^ r = l i k i Trace ( L r ) = x i k i f ( x k i ) Trace ( L r )
By conditional relationship formula (5):
Figure BDA0000420595640000057
and formula (4) has
x i k i f ( x k i ) > x i k j f ( x k j )
Be multiplied by simultaneously on above formula both sides
Figure BDA0000420595640000059
divided by Trace (L r):
1 Trace ( L r ) x j k j x i k i f ( x k i ) > 1 Trace ( L r ) x j k j f ( x k j ) x i k j
Above formula is equivalent to (10), if institute the leaf node of looking for for
Figure BDA00004205956400000511
meet formula 10), this leaf node district inclusion is vectorial
Figure BDA00004205956400000512
Technical conceive of the present invention is: with two frontier points, start to set up support vector, form initial support matrix, take this initial matrix to set up binary tree as root; Then known data point is set up to support vector, and upgrade this binary tree according to condition; A sampling step length is set again, within the scope of the field of definition of x, samples, find out the affiliated leaf node of sampled point, calculate the low valuation of lower bound of the affiliated leaf node region of sampled point; Connect the lower bound sawtooth estimation that all sampled points just can obtain treating matched curve; With same method, try to achieve the upper bound sawtooth estimation for the treatment of matched curve, the curve on average obtaining of getting the upper bound and the estimation of lower bound sawtooth is treats matched curve.The abstract convex estimation curve matching of application based on tree, has not only improved efficiency and the accuracy rate of matching, and has retained raw data points, has reduced calculated amount.
Accompanying drawing explanation
Fig. 1 is the unit's of being converted to simplex space, common definition territory schematic diagram, and wherein, left side subgraph is common definition territory, and the right subgraph is unit simplex space.
Fig. 2 is the schematic diagram that the lower bound sawtooth of one dimension Shubert function is estimated.
Fig. 3 is the schematic diagram that one dimension Shubert function upper bound sawtooth is estimated.
Fig. 4 is the schematic diagram of the matched curve of one dimension Shubert function.
Embodiment
Below in conjunction with accompanying drawing, the invention will be further described.
With reference to Fig. 1~Fig. 4, a kind of curve-fitting method of estimating based on abstract convex, comprises the following steps:
1) parameter initialization: constant M is set, input data (x k, y k) (k=1,2 ..., K), the quantity size size that wherein K is data point.
2) boundary condition initialization.
2.1) according to existing data point, determine the field of definition [a, b] of variable x, wherein a is the lower bound of variable x, and b is the upper bound;
2.2) model conversion;
If x is ∈ [a, b], wherein, x is the variable on boundary constraint feasible zone [a, b], and former variable x is done to following linear transformation:
Figure BDA0000420595640000061
X wherein 1', x 2' be linear transformation variable.
Obviously, (b-a) > 0, and x is easily known in (x-a)>=0 1'>=0, x 2'>=0, x ′ = [ x 1 ′ , x 2 ′ ] T ∈ S = { x ′ ∈ R + 2 : Σ i = 1 2 x i ′ = 1 } , Wherein R + 2 = { x ′ = ( x 1 ′ , x 2 ′ ) | x i ′ ≥ 0 , i = 1,2 } , S representation unit simplex region; Fig. 1 has provided the schematic diagram in the unit's of being converted to simplex space, common definition territory.
Formula (1) is converted, can obtain:
x=x 1′(b-a)+a (2)
2.2.1) establish
Figure BDA0000420595640000064
wherein,
Figure BDA0000420595640000065
the linear transformation variable 1 that represents the 1st data,
Figure BDA0000420595640000066
the linear transformation variable 1 of the 2nd data;
2.2.2) right according to formula (2)
Figure BDA0000420595640000071
be converted to former variable x 1, x 2, x wherein 1=b, x 2=a;
2.3) calculate initial support vector sum proppant matrix, and set up tree;
Definition 1: if function f meets following two conditions
i ) ∀ x , y ∈ R + n , x ≥ y ⇒ f ( x ) ≥ f ( y )
ii ) ∀ x ∈ R + n , ∀ λ ∈ R + + : f ( λx ) = λf ( x )
Wherein,
Figure BDA00004205956400000714
represent arbitrarily,
Figure BDA0000420595640000074
claim function f for just homogeneous increasing function (Increasing Positively Homogeneous of degree one, IPH).
Objective function also can by add one enough large constant M be converted to IPH function.Suppose that the field of definition of N dimension objective function utilized model conversion formula (1) unit's of being mapped to simplex space
Figure BDA0000420595640000076
suppose that it has K support function, the point (x based on given k, f (x k)) (k=1,2 ..., K), the Lower Bound Estimation of objective function f can be provided by following formula:
H K ( x ) = max k ≤ K min i = 1 , . . . N + 1 x i l i k - - - ( 3 )
Wherein max represents maximum, and min represents minimum, x ifor the vector in unit simplex space;
K support function is:
h k ( x ) = min i = 1 , . . . N + 1 x i l i k = f ( x k ) min { x 1 x 1 k , . . . , x N + 1 x N + 1 k }
Wherein
l k = ( x 1 k f ( x k ) , x 2 k f ( x k ) , . . . , x N + 1 k f ( x k ) ) - - - ( 4 )
Be called support vector.Consider a set that contains K support vector
Figure BDA00004205956400000710
make I={1,2 ..., N+1}, N+1 support vector of two conditions (5) and (6) below meeting
Figure BDA00004205956400000711
be called an effective proppant matrix;
∀ i , j ∈ I , i ≠ j : l i k i > l i k j - - - ( 5 )
∀ v ∈ Λ k \ L , ∃ i ∈ I : l i k i ≤ v i - - - ( 6 )
V ∈ Λ wherein kl represent that v belongs to Λ kbut do not belong to L,
Figure BDA00004205956400000715
represent to exist;
Explain for convenience principle, suppose an effective proppant matrix as the formula (7):
Figure BDA0000420595640000081
According to this proppant matrix, can calculate corresponding lower bound support function extreme value x minthe Lower Bound Estimation value d corresponding with it:
x min(L)=diag(L)/Trace(L) (8)
d(L)=H k(x min)=1/Trace(L) -1 (9)
Wherein, diag is the positive diagonal entry of matrix, and Trace is matrix trace, i.e. the positive diagonal entry sum of matrix.
2.3.1) according to formula (4) respectively calculation level (x ' 1, y b) and (x ' 2, y a) support vector, wherein, y brepresent functional value corresponding to upper bound b, y arepresent functional value corresponding to lower bound a, the x in formula kwith x ' replacement, f () replaces with y;
2.3.2) 2 support vectors of gained are formed to proppant matrix L, as shown in Equation (7);
2.3.3) take proppant matrix L as root foundation tree;
3) calculate the support vector of each data point and upgrade tree, each data point is handled as follows:
3.1) according to formula (1) to x kas linear transformation, be
3.2) according to formula (4) calculate point (x ' k, y k) support vector l k, the x in formula kwith x ' kreplace, f () uses y kreplace;
3.3) according to conditional relationship formula (5) and (6), upgrade tree;
3.3.1) find for step 3.2) the support vector l that produces k(be v i=l k) leaf node of do not satisfy condition (6);
3.3.2) use l kreplacement step 3.3.1) i support vector in the leaf node matrix finding in
Figure BDA0000420595640000083
thereby form new leaf node;
3.3.3) the new leaf node producing the determining step 3.3.2) relational expression (5) that whether satisfies condition, if met, retains, otherwise deletes;
4) calculate Lower Bound Estimation value.
Suppose
Figure BDA0000420595640000084
for H k(x) r local minimum, the relational expression that satisfies condition (5) of its correspondence and the proppant matrix of (6) are L r.By formula (9), had
Figure BDA0000420595640000085
lower Bound Estimation value
H k ( x ^ r ) = 1 / Trace ( L r ) - 1
i element be
x ^ r = l i k i Trace ( L r ) = x i k i f ( x k i ) Trace ( L r )
By conditional relationship formula (5):
Figure BDA0000420595640000093
and formula (4) has
x i k i f ( x k i ) > x i k j f ( x k j )
Be multiplied by simultaneously on above formula both sides divided by Trace (L r):
1 Trace ( L r ) x j k j x i k i f ( x k i ) > 1 Trace ( L r ) x j k j f ( x k j ) x i k j
Above formula is equivalent to:
x j k j x ^ i r > x i k j x ^ j r , i , j ∈ I , i ≠ j - - - ( 10 )
Wherein
Figure BDA0000420595640000098
for the element in looked for leaf node matrix.
4.1) sampling step length is set, and sampling in field of definition [a, b] scope;
4.2) to step 4.1) each sampled point x cbe handled as follows:
4.2.1) according to formula (1), do linear transformation, obtain new vector x c'=(x c1', x c2');
4.2.2) according to formula (10) from step 3.3) tree find and comprise vector x c' leaf node, in formula
Figure BDA0000420595640000099
use x c' replace, if leaf node meets formula (10), comprise, otherwise do not comprise;
4.2.3) according to formula (3), obtain vector x cthe Lower Bound Estimation value y of the corresponding subregion of leaf node at ' place l';
4.2.5) due in step 2.3) in added a constant M while being converted into IPH, in the hope of Lower Bound Estimation value should deduct M, i.e. y l=y l'-M;
4.2.6) output Lower Bound Estimation value y land corresponding sampled point x c;
5) calculate Estimation of Upper-Bound value.
5.1) the y in data point kget negative;
5.2) by step 2) to 4) calculating Lower Bound Estimation value;
5.3) to step 5.2) the Lower Bound Estimation value of gained gets negatively, is Estimation of Upper-Bound value y u;
5.4) output Estimation of Upper-Bound value y uand corresponding sampled point x c;
6) curve.
6.1) connect each point (x c, y l) obtain treating the Lower Bound Estimation of matched curve;
6.2) connect each point (x c, y u) obtain treating the Estimation of Upper-Bound of matched curve;
6.3) Estimation of Upper-Bound and Lower Bound Estimation being averaged to the curve obtaining is and treats matched curve.
It is embodiment that the present embodiment be take one dimension Shubert function, and a kind of curve-fitting method of estimating based on abstract convex, wherein comprises following steps:
1) parameter initialization: constant M=300 is set, 100 data point (x of random generation in field of definition [0,5] k, y k), k=1,2 ..., 100.
2) boundary condition initialization.
2.1) field of definition of determining variable x is [0,5];
2.2) model conversion;
If x is ∈ [a, b], wherein, x is the variable on boundary constraint feasible zone [a, b], and former variable x is done to following linear transformation:
Figure BDA0000420595640000101
X wherein 1', x 2' be linear transformation variable.
Obviously, (b-a) > 0, and x is easily known in (x-a)>=0 1'>=0, x 2'>=0, x ′ = [ x 1 ′ , x 2 ′ ] T ∈ S = { x ′ ∈ R + 2 : Σ i = 1 2 x i ′ = 1 } , Wherein R + 2 = { x ′ = ( x 1 ′ , x 2 ′ ) | x i ′ ≥ 0 , i = 1,2 } , S representation unit simplex region; Fig. 1 has provided the schematic diagram in the unit's of being converted to simplex space, common definition territory.
Formula (1) is converted, can obtain:
x=x 1′(b-a)+a (2)
2.2.1) establish
Figure BDA0000420595640000104
wherein,
Figure BDA0000420595640000105
the linear transformation variable 1 that represents the 1st data,
Figure BDA0000420595640000106
the linear transformation variable 1 of the 2nd data;
2.2.2) right according to formula (2)
Figure BDA0000420595640000107
be converted to former variable x 1, x 2, x wherein 1=5, x 2=0;
2.3) calculate initial support vector sum proppant matrix, and set up tree;
Definition 1: if function f meets following two conditions
i ) ∀ x , y ∈ R + n , x ≥ y ⇒ f ( x ) ≥ f ( y )
ii ) ∀ x ∈ R + n , ∀ λ ∈ R + + : f ( λx ) = λf ( x )
Wherein,
Figure BDA00004205956400001012
represent arbitrarily,
Figure BDA00004205956400001010
claim function f for just homogeneous increasing function (Increasing Positively Homogeneous of degree one, IPH).
Objective function also can add one enough large constant M be converted to IPH function.Suppose that the field of definition of N dimension objective function utilized model conversion formula (1) unit's of being mapped to simplex space
Figure BDA0000420595640000111
suppose that it has K support function, the point (x based on given k, f (x k)) (k=1,2 ..., K), the Lower Bound Estimation of objective function f can be provided by following formula:
H K ( x ) = max k ≤ K min i = 1 , . . . N + 1 x i l i k - - - ( 3 )
Wherein max represents maximum, and min represents minimum, x ifor the vector in unit simplex space;
K support function is:
h k ( x ) = min i = 1 , . . . N + 1 x i l i k = f ( x k ) min { x 1 x 1 k , . . . , x N + 1 x N + 1 k }
Wherein
l k = ( x 1 k f ( x k ) , x 2 k f ( x k ) , . . . , x N + 1 k f ( x k ) ) - - - ( 4 )
Be called support vector.Consider a set that contains K support vector
Figure BDA0000420595640000115
make I={1,2 ..., N+1}, N+1 support vector of two conditions (5) and (6) below meeting
Figure BDA0000420595640000116
be called an effective proppant matrix.
∀ i , j ∈ I , i ≠ j : l i k i > l i k j - - - ( 5 )
∀ v ∈ Λ k \ L , ∃ i ∈ I : l i k i ≤ v i - - - ( 6 )
Explain for convenience principle, suppose an effective proppant matrix as the formula (7):
Figure BDA0000420595640000119
According to this proppant matrix, can calculate corresponding lower bound support function extreme value x minthe Lower Bound Estimation value d corresponding with it:
x min(L)=diag(L)/Trace(L) (8)
d(L)=H k(x min)=1/Trace(L) -1 (9)
Wherein, diag is the positive diagonal entry of matrix, and Trace is matrix trace, i.e. the positive diagonal entry sum of matrix.
2.3.1) according to formula (4) respectively calculation level (x ' 1, y 5) and (x ' 2, y 0) support vector, wherein, y 5the functional value that represents the upper bound 5 correspondences, y 0the functional value that represents lower bound 0 correspondence, the x in formula kwith x ' replacement, f () replaces with y;
2.3.2) 2 support vectors of gained are formed to proppant matrix L, as shown in Equation (7);
2.3.3) take proppant matrix L as root foundation tree;
3) calculate the support vector of each data point and upgrade tree, each data point is handled as follows:
3.1) according to formula (1) to x kas linear transformation, be
Figure BDA0000420595640000121
3.2) according to formula (4) calculate point (x ' k, y k) support vector l k, the x in formula kwith x ' kreplace, f () uses y kreplace;
3.3) according to conditional relationship formula (5) and (6), upgrade tree;
3.3.1) find for step 3.2) the support vector l that produces k(be v i=l k) leaf node of do not satisfy condition (6);
3.3.2) use l kreplacement step 3.3.1) i support vector in the leaf node matrix finding in
Figure BDA0000420595640000122
thereby form new leaf node;
3.3.3) the new leaf node producing the determining step 3.3.2) relational expression (5) that whether satisfies condition, if met, retains, otherwise deletes;
4) calculate Lower Bound Estimation value.
Suppose
Figure BDA0000420595640000123
for H k(x) r local minimum, the relational expression that satisfies condition (5) of its correspondence and the proppant matrix of (6) are L r.By formula (9), had
Figure BDA0000420595640000124
lower Bound Estimation value
H k ( x ^ r ) = 1 / Trace ( L r ) - 1
Figure BDA0000420595640000126
i element be
x ^ r = l i k i Trace ( L r ) = x i k i f ( x k i ) Trace ( L r )
By conditional relationship formula (5): and formula (4) has
x i k i f ( x k i ) > x i k j f ( x k j )
Be multiplied by simultaneously on above formula both sides
Figure BDA00004205956400001210
divided by Trace (L r):
1 Trace ( L r ) x j k j x i k i f ( x k i ) > 1 Trace ( L r ) x j k j f ( x k j ) x i k j
Above formula is equivalent to:
x j k j x ^ i r > x i k j x ^ j r , i , j ∈ I , i ≠ j - - - ( 10 )
Wherein
Figure BDA00004205956400001213
for the element in looked for leaf node matrix.
4.1) sampling step length being set is 0.001, and sampling in field of definition [0,5] scope;
4.2) to step 4.1) each sampled point x cbe handled as follows:
4.2.1) according to formula (1), do linear transformation, obtain new vector x c'=(x c1', x c2');
4.2.2) according to formula (10) from step 3.3) tree find and comprise vector x c' leaf node, in formula use x c' replace, if leaf node meets formula (10), comprise, otherwise do not comprise;
4.2.3) according to formula (3), obtain vector x cthe Lower Bound Estimation value y of the leaf node corresponding region at ' place l';
4.2.5) due in step 2.3) in added a constant M while being converted into IPH, in the hope of Lower Bound Estimation value should deduct M, i.e. y l=y l'-M;
4.2.6) output Lower Bound Estimation value y land corresponding sampled point x c;
5) calculate Estimation of Upper-Bound value.
5.1) the y in data point kget negative;
5.2) by step 2) to 4) calculating Lower Bound Estimation value;
5.3) to step 5.2) the Lower Bound Estimation value of gained gets negatively, is Estimation of Upper-Bound value y u;
5.4) output Estimation of Upper-Bound value y uand corresponding sampled point x c;
6) curve.
6.1) connect each point (x c, y l) the lower bound sawtooth that obtains one dimension Shubert function curve estimates, as shown in Figure 2;
6.2) connect each point (x c, y u) upper bound sawtooth that obtains one dimension Shubert function curve estimates, as shown in Figure 3;
6.3) Estimation of Upper-Bound and Lower Bound Estimation are averaged to the matched curve that the curve obtaining is one dimension Shubert function, as shown in Figure 4.
The one dimension Shubert function of take is embodiment, uses above method to obtain its matched curve as shown in Figure 4, and in figure, " * " point is known data point, and sawtooth curve is the matched curve of one dimension Shubert function.As shown in Figure 2, in figure, " * " point is known data point in the lower bound sawtooth estimation of one dimension Shubert function, and jagged equity curve is that the lower bound sawtooth of one dimension Shubert function is estimated.As shown in Figure 3, in figure, " * " point is known data point in upper bound sawtooth estimation, and jagged equity curve is that the upper bound sawtooth of one dimension Shubert function is estimated.
What more than set forth is the excellent results that embodiment shows that the present invention provides, obviously the present invention is not only applicable to above-described embodiment, can do many variations to it and is implemented not departing from essence spirit of the present invention and do not exceed under the prerequisite of the related content of flesh and blood of the present invention.

Claims (4)

1. a curve-fitting method of estimating based on abstract convex, is characterized in that: described curve-fitting method comprises the following steps:
1) parameter initialization: constant M is set, input data (x k, y k), k=1,2 ..., K, the quantity size size that wherein K is data point;
2) boundary condition initialization, process is as follows;
2.1) according to existing data point, determine the field of definition [a, b] of variable x, wherein a is the lower bound of variable x, and b is the upper bound;
2.2) model conversion: former variable x is done to following linear transformation:
X wherein 1', x 2' be linear transformation variable;
Formula (1) is converted, obtains:
x=x 1′(b-a)+a (2)
2.2.1) establish wherein,
Figure RE-FDA0000470342390000015
the linear transformation variable 1 that represents the 1st data,
Figure RE-FDA0000470342390000016
the linear transformation variable 1 of the 2nd data;
2.2.2) right according to formula (2) be converted to former variable x 1, x 2, x wherein 1=b, x 2=a;
2.3) calculate initial support vector sum proppant matrix, and set up tree:
2.3.1) according to formula (4) respectively calculation level (x ' 1, y b) and (x ' 2, y a) support vector,
Figure RE-FDA0000470342390000012
Wherein, y brepresent functional value corresponding to upper bound b, y arepresent functional value corresponding to lower bound a, the x in formula kwith x ' replacement, f () replaces with y;
2.3.2) two support vectors of gained are formed to proppant matrix L, as shown in Equation (7):
Figure RE-FDA0000470342390000013
2.3.3) take proppant matrix L as root foundation tree;
3) calculate the support vector of each data point and upgrade tree, each data point is handled as follows:
3.1) according to formula (1) to x kas linear transformation, be
Figure RE-FDA0000470342390000027
3.2) according to formula (4) calculate point (x ' k, y k) support vector l k, the x in formula kwith x ' kreplace, f () uses y kreplace;
3.3) according to conditional relationship formula (5) and (6), upgrade tree:
Figure RE-FDA0000470342390000021
Figure RE-FDA0000470342390000022
V ∈ Λ wherein kl represent that v belongs to Λ kbut do not belong to L, represent to exist;
3.3.1) find for step 3.2) the support vector l that produces kthe leaf node of (6), i.e. v do not satisfy condition i=l k;
3.3.2) use l kreplacement step 3.3.1) i support vector in the leaf node matrix finding in
Figure RE-FDA0000470342390000023
thereby form new leaf node;
3.3.3) the new leaf node producing the determining step 3.3.2) relational expression (5) that whether satisfies condition, if met, retains, otherwise deletes;
4) calculate Lower Bound Estimation value;
4.1) sampling step length is set, and sampling in field of definition [a, b] scope;
4.2) to step 4.1) each sampled point x cbe handled as follows:
4.2.1) according to formula (1), do linear transformation, obtain new vector x c'=(x c' 1, x c' 2);
4.2.2) according to formula (10) from step 3.3) tree find and comprise vector x c' leaf node, in formula
Figure RE-FDA0000470342390000024
use x c' replace, if leaf node meets formula (10), comprise, otherwise do not comprise;
Wherein, for the element in looked for leaf node matrix;
4.2.3) according to formula (3), obtain vector x cthe Lower Bound Estimation value y of the leaf node corresponding region at ' place l';
4.2.4) due in step 2.3) in added a constant M while being converted into IPH, in the hope of Lower Bound Estimation value should deduct M, i.e. y l=y l'-M;
4.2.5) output Lower Bound Estimation value y land corresponding sampled point x c;
5) calculate Estimation of Upper-Bound value, process is as follows:
5.1) the y in data point kget negative;
5.2) by step 2) to 4) calculating Lower Bound Estimation value;
5.3) to step 5.2) the Lower Bound Estimation value of gained gets negatively, is Estimation of Upper-Bound value y u;
5.4) output Estimation of Upper-Bound value y uand corresponding sampled point x c;
6) curve, process is as follows:
6.1) connect each point (x c, y l) obtain treating the Lower Bound Estimation of matched curve;
6.2) connect each point (x c, y u) obtain treating the Estimation of Upper-Bound of matched curve;
6.3) Estimation of Upper-Bound and Lower Bound Estimation being averaged to the curve obtaining is and treats matched curve.
2. a kind of curve-fitting method of estimating based on abstract convex as claimed in claim 1, is characterized in that: the constant M in step 1), if function f meets following two conditions:
Figure FDA0000420595630000031
Figure FDA0000420595630000032
Wherein,
Figure FDA0000420595630000038
represent arbitrarily,
Figure FDA0000420595630000033
Figure FDA0000420595630000034
claim function f for just homogeneous increasing function IPH;
Add one enough large constant M can will treat that function corresponding to matched curve is converted to IPH function, adds M to each y in data.
3. a kind of curve-fitting method of estimating based on abstract convex as claimed in claim 1 or 2, is characterized in that: step 2.3) in calculate support vector and proppant matrix method be:
Suppose that the field of definition of N dimension objective function utilized model conversion formula (1) unit's of being mapped to simplex space
Figure FDA0000420595630000035
suppose that it has K support function, the point (x based on given k, f (x k)) (k=1,2 ..., K), the Lower Bound Estimation of objective function f is provided by following formula:
Figure FDA0000420595630000036
Wherein max represents maximum, and min represents minimum, x ifor the vector in unit simplex space;
K support function is:
Figure FDA0000420595630000037
Wherein, l kbe called support vector, as the formula (4), consider a set that contains K support vector
Figure FDA0000420595630000041
make I={1,2 ..., N+1}, N+1 the support vector of satisfy condition (5) and (6)
Figure FDA0000420595630000042
be called an effective proppant matrix;
Suppose an effective proppant matrix as the formula (7):
Figure FDA0000420595630000043
According to this proppant matrix, can calculate corresponding lower bound support function extreme value x minthe Lower Bound Estimation value d corresponding with it:
x min(L)=diag(L)/Trace(L) (8)
d(L)=H k(x min)=1/Trace(L) -1 (9)
Wherein, diag is the positive diagonal entry of matrix, and Trace is matrix trace, i.e. the positive diagonal entry sum of matrix.
4. a kind of curve-fitting method of estimating based on abstract convex as claimed in claim 3, is characterized in that: step 4.2.2), find out and comprise vector x c' the method for leaf node be:
Suppose
Figure FDA0000420595630000044
for H k(x) r local minimum, the relational expression that satisfies condition (5) of its correspondence and the proppant matrix of (6) are L r, by formula (9), had lower Bound Estimation value
Figure FDA0000420595630000046
Figure FDA0000420595630000047
i element be
Figure FDA0000420595630000048
By conditional relationship formula (5):
Figure FDA0000420595630000049
and formula (4) has
Figure FDA00004205956300000410
Be multiplied by simultaneously on above formula both sides
Figure FDA00004205956300000411
divided by Trace (L r):
Figure FDA00004205956300000412
Above formula is equivalent to formula (10), if institute the leaf node of looking for for
Figure FDA00004205956300000413
meet formula 10), this leaf node district inclusion is vectorial
Figure FDA00004205956300000414
CN201310600939.7A 2013-11-22 2013-11-22 Curve fitting method based on abstract convex estimations Pending CN103745079A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201310600939.7A CN103745079A (en) 2013-11-22 2013-11-22 Curve fitting method based on abstract convex estimations

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201310600939.7A CN103745079A (en) 2013-11-22 2013-11-22 Curve fitting method based on abstract convex estimations

Publications (1)

Publication Number Publication Date
CN103745079A true CN103745079A (en) 2014-04-23

Family

ID=50502096

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201310600939.7A Pending CN103745079A (en) 2013-11-22 2013-11-22 Curve fitting method based on abstract convex estimations

Country Status (1)

Country Link
CN (1) CN103745079A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104182653A (en) * 2014-08-25 2014-12-03 浙江工业大学 Group conformation space optimizing method on basis of local abstract convexity underestimated support surface
WO2018076211A1 (en) * 2016-10-26 2018-05-03 中国科学院自动化研究所 Method for quadratic curve fitting in image based on geometric error optimization
WO2023165268A1 (en) * 2022-03-04 2023-09-07 支付宝(杭州)信息技术有限公司 Data processing

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104182653A (en) * 2014-08-25 2014-12-03 浙江工业大学 Group conformation space optimizing method on basis of local abstract convexity underestimated support surface
WO2018076211A1 (en) * 2016-10-26 2018-05-03 中国科学院自动化研究所 Method for quadratic curve fitting in image based on geometric error optimization
WO2023165268A1 (en) * 2022-03-04 2023-09-07 支付宝(杭州)信息技术有限公司 Data processing

Similar Documents

Publication Publication Date Title
Kisi Pan evaporation modeling using least square support vector machine, multivariate adaptive regression splines and M5 model tree
Liu et al. Short-term wind speed forecasting based on spectral clustering and optimised echo state networks
Sedki et al. Hybrid particle swarm optimization and differential evolution for optimal design of water distribution systems
US20170338802A1 (en) Actually-measured marine environment data assimilation method based on sequence recursive filtering three-dimensional variation
Ji et al. Predicting dynamic deformation of retaining structure by LSSVR-based time series method
CN110659693B (en) K-nearest neighbor classification-based power distribution network rapid topology identification method, system and medium
Afshar et al. Optimizing water supply and hydropower reservoir operation rule curves: an imperialist competitive algorithm approach
CN105225000B (en) A kind of wind power probability model nonparametric probability method based on Fuzzy Ordered optimization
CN103020709B (en) Based on the one-dimensional water quality model parameter calibration method of artificial bee colony and quanta particle swarm optimization
CN107274028A (en) A kind of many wind fields based on mixing Copula functions are exerted oneself Forecasting Methodology
CN109558968B (en) Wind farm output correlation analysis method and device
CN103116698A (en) GM (1, 1) model prediction method based on cubic spline
CN103745079A (en) Curve fitting method based on abstract convex estimations
Qian et al. An ant colony system based on the Physarum network
CN104156768A (en) Small data size chaos identifying method through fuzzy C-means cluster
Sayyaadi et al. Stability and iterative convergence of water cycle algorithm for computationally expensive and combinatorial Internet shopping optimisation problems
Akbari et al. Similarity-based error prediction approach for real-time inflow forecasting
Yang et al. Forecasting time series with genetic programming based on least square method
Li et al. Uncertain and multi-objective programming models for crop planting structure optimization
CN104732292A (en) Automatic planning method for main velocity line based on profile data
CN104850646A (en) Method of mining frequent subgraphs for single uncertain graphs
Yu et al. Forecasting model of agricultural products prices in wholesale markets based on combined BP neural network-time series model
CN104881525A (en) 3D (three dimension) surface fitting method based on Lipschitz estimation
Fatholahi et al. Development gap in Provinces of Iran
Prasad et al. A stability analysis of logistic model

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20140423

RJ01 Rejection of invention patent application after publication