Apparatus for and method of generating an approximation function

• US 5,519,647 A
• Filed: 05/11/1994
• Issued: 05/21/1996
• Est. Priority Date: 05/12/1993
• Status: Expired due to Fees

• Current Assignee: Adelante Technologies B.V.
• Original Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Inventors: DeVille, Yannick
• Primary Examiner(s): Envall, Jr., Roy N.
• Assistant Examiner(s): MOISE, EMMANUEL LIONEL
• Application Number:
  US08/241,088
• Time in Patent Office: 741 Days
• Field of Search: 364/721, 364/718, 364/719, 364/720, 364/723, 364/735, 395/21
• US Class Current: 708/276
• CPC Class Codes: G06F 17/17 Function evaluation by approximation methods, e.g. inter- or extrapolation, smoothing, least mean square method G06F17/18 takes precedence ; interpolation for numerical control G05B19/18
• CPC Class Codes: G06F 17/18 for evaluating statistical data , e.g. average values, frequency distributions, probability functions, regression analysis forecasting specially adapted for a specific administrative, business or logistic context G06Q10/04

Abstract

An apparatus (5) for generating an approximation function based on first pairs ((X₁, Y₁) to (X₆, Y₆)) of values associating a dependent variable (Y₁ to Y₆) with an independent variable (X₁ to X₆), and for determining second pairs (X_A, Y'"'"'_A) of values of said variables in accordance with said approximation function. The apparatus comprises: a) first means (10) for iteratively determining at least one current linear regression function, for selecting that one of the current linear functions which produces the approximation of all the pairs of said series with minimal errors, and for coding the selected linear regression function with the aid of specific codes (p, q), and b) second means (17) for determining said second pairs (X_A, Y'"'"'_A) with the aid of said specific codes. The apparatus can also be used for calculating approximated values of mathematical functions, for example a in a neural network, or for determining a regression function forming an approximation to experimental measurement results, for example distributed measurements resulting from monitoring an industrial process. The invention also relates to a method of generating an approximation function.

Claims

1. An apparatus (5) for generating an approximation function based on first pairs ((X₁, Y₁) to (X₆, Y₆)) of values associating a dependent variable (Y₁ to Y₆) with an independent variable (X₁ to X₆), and for determining second pairs (X_A, Y'"'"'_A) of values of said variables in accordance with said approximation function, characterized in that the apparatus comprises:

first means (10);

for iteratively determining at least one current linear regression function by making first errors of alternating sign equal in absolute value, which first errors have been measured between, respectively, first values (Y₃, Y₄, Y₅) of the dependent variable for three pairs (X₃, Y₃) (X₄, Y₄) (X₅, Y₅) of a series of pairs, and second values (Y'"'"'₃, Y'"'"'₄, Y'"'"'₅) of the dependent variable determined, in accordance with said current linear regression function, for the same values (X₃, X₄, X₅) of the independent variable,

for selecting that one of the current linear regression functions which produces an approximation of all the pairs of said series with a minimum error, and

for coding the selected linear regression function with the aid of specific codes (p, q), and

second means (17) for determining said second pairs (X_A, Y'"'"'_A) with the aid of said specific codes.

2. An apparatus as claimed in claim 1, characterized in that the first means (10) comprise:

storage means (12c) for storing the values of the first pairs and said specific codes,

calculation means (13c) (19c) for successively calculating the specific codes (p, q) of current linear regression functions and their respective first errors, equalized in absolute value (E_T), and for selecting at least one current linear regression function, and

means (11c) for controlling the storage means (12c) and the calculation means (13c).

3. An apparatus (5) as claimed in claim 1, characterized in that for selecting said linear regression function the first means comprise means (14c):

for likewise determining additional errors for the other pairs (X₁, Y₁) (X₂, Y₂) (X₆, Y₆) of values of said series,

for comparing the additional errors with said equalized first errors (E_T), and

for selecting the current linear regression function which yields additional errors which, in absolute value, are smaller than or equal to said equalized first errors (E_T).

4. An apparatus as claimed in claim 2, characterized in that for selecting the linear regression function the calculation means (13c) select that one of the current linear regression functions which yields maximal equalized first errors (E_T).

5. An apparatus as claimed in any one of the claim 1, characterized in that said second means (17) comprise:

storage means (12a) for storing said specific codes of at least one linear regression function,

calculation means (13a) for selecting, in response to a value of the independent variable (X_A), a linear regression function belonging to said value, and for calculating, starting from the specific codes received from the storage means, a value of the dependent variable (Y'"'"'_A) associated with the value of the independent variable (X_A), and

means (11a) for addressing the specific codes of the linear regression function to the calculation means at the request of said means.

6. In an apparatus for generating an approximation function based on first pairs ((X₁, Y₁) to (X₆, Y₆)) of values associating a dependent variable (Y₁ to Y₆) with an independent variable (X₁ to X₆), a method of generating said approximation function based on first pairs (X₁, Y₁) (X₆, Y₆) of values, and for determining second pairs (X_A, Y'"'"'_A) of values of said variables in accordance with said approximation function, characterized in that the method comprises:

a first phase;

for iteratively determining at least one current linear regression function by making first errors (E_PD) of alternating sign equal in absolute value, which first errors have been measured between, respectively, first values (Y₃, Y₄, Y₅) of the dependent variable for three pairs (X₃, Y₃) (X₄, Y₄) (X₅, Y₅) of a series of pairs, and second values (Y'"'"'₃, Y'"'"'₄, Y'"'"'₅) of the dependent variable determined, in accordance with said current linear regression function, for the same values (X₃, X₄ 4, X₅) of the independent variable,

for selecting that one of the current linear regression functions which produces an approximation of all the pairs of said series with a minimum error, and

for coding the selected linear regression function with the aid of specific codes, and

second means (17) for determining said second pairs (X_A, Y'"'"'_A) with the aid of said specific codes.

7. A method as claimed in claim 6, characterized in that the first phase includes the following steps:

A--selecting (102) three pairs (P3, P4, P5) of values from said series,

B--calculating (104) the current linear regression function D and determining an associated triplet error E_T =|E_PD |,

C--selecting (106) an additional pair (P1), (P2), (P6),

D--calculating (108) an additional error E_Pm between the additional pair and said function,

E--if |E_Pm |≦

E_T (110) for the additional pair (110) the method proceeds with step C with a following additional pair, and

F--if |E_Pm |>

E_T for at least one additional pair (110) the method proceeds with step A with a selection of another group of three pairs from said series of pairs, and

G--if |E_Pm |≦

E_T for all the additional pairs (112) the current linear regression function is coded and stored (114) as a linear approximation function.

8. A method as claimed in claim 6, characterized in that the first phase includes the following steps:

A--selecting (102) three pairs (P3, P4, P5) of values from said series,

B--calculating (104) the current linear regression function D and determining an associated triplet error E_T =|E_PD |,

C--comparing (210) the error E_T with an optimum error E_op having a strictly negative initial value,

D--if E_T >

E_op updating (114) the optimum error E_op by replacing E_op by E_T and updating the codes of an optimum linear regression function D_op by replacing these codes by the codes of the current linear regression function (D),

E--subsequently returning (116) to step A to select three other pairs, and

F--when all the triplets of pairs of values of the series have been checked (116) the last codes of the optimum linear regression function D_op will form the codes of the linear approximation function (114).

9. A method as claimed in claim 8, characterized in that the first phase includes the following modified steps:

A1--modification of step A by effecting a selection (102a) of two pairs of values belonging to said series, such that there is at least one additional intermediate pair with an independent variable (X₁ -X₆) situated between the independent variables of said pair in order to form at least one triplet of pairs of values,

A2--modification of step A, firstly, by determining (102b) a subsidiary linear function which contains the two selected pairs and, secondly, by determining second errors between the dependent variables of the possible intermediate pairs and said subsidiary linear function;

and, when these second errors all have the same sign, selecting the intermediate pair producing the second error having the largest absolute value in order to form a triplet of pairs of values formed by the intermediate pair and the two selected pairs,

and, when these second errors have different signs, resuming the method in step A1,

B1--step B (104) is effected with said selected triplet,

C1--modification of step C (106) by selecting an additional pair whose independent variable is not situated between the independent variables of the two selected pairs.

10. A method as claimed in claim 9, modified in that, prior to the step A1, said pairs of values being represented as points (P1-P6) in a two-dimensional space, the first phase of the method includes a step (100a) to determine a lower envelope and/or an upper envelope which interconnect the outer points of the set of points, the selection of the two pairs of values in the step A1 being made from adjacent points belonging to one of the envelopes.

11. A method as claimed in in any one of the claim 6, characterized in that a specific weighting coefficient, which weights said first errors, is associated with each value in the series of values of the independent variable.

12. A method as claimed in any one of the claim 6, characterized in that it generates an approximation function comprising a series of interlinked linear regression functions.

13. Data processing system comprising, in succession, an input neuron layer, at least one hidden neuron layer, and an output neuron layer, means for calculating neural potentials of the neurons in the output neuron layer by cumulating activities of previous layers connected thereto, means for delivering output neurons states by applying an activation function to the neural potentials, wherein the system comprises:

a) means for storing pairs of values of the activation function, as values of an independent variable and as values of a dependent variable defining the activation function,

b) first calculation means for computing a linear regression function approximating the pairs of values within a predetermined error, and for coding the linear regression function,

c) second calculation means addressing the codes for computing the output neuron states by applying the activation function to the neural potentials on the basis of the codes of the linear regression function.

14. Data processing system as claimed in claim 13 wherein the first calculation means determines the linear regression function by

equalizing, in absolute value, a primary error over two or three pairs of values and

testing whether secondary errors for the remaining pairs of values are less than or equal to the primary error,

the primary and secondary errors being measured between each pair of values and the linear regression function.

15. Data processing system as claimed in claim 14 wherein the primary error and the secondary error are weighted with specific weighting coefficients.

16. Data processing system as claimed in claim 14 wherein the first calculation means splits the pairs of values in a plurality of sub-sets, the first calculation means determining one linear regression function for each sub-set and computing the codes for each sub-set, the system comprising means for selecting the sub-set specific to the output neuron potential to process and for addressing to the second calculation means the codes peculiar to the linear regression function of the sub-set selected.

17. Apparatus for generating an approximation function approximating first pairs of values, a pair comprising one value of an independent variable and one value of a dependent variable, and for computing second pairs of values based on the approximation function for request values of the independent variable, wherein the apparatus comprises:

a) means for storing the first pairs of values,

b) first calculation means for computing the approximation function as a linear regression function approximating the first pairs of values to within a predetermined error, and for delivering codes coding the linear regression function,

c) second calculation means for computing the second pairs of values based on the codes of the linear regression function.

18. Apparatus as claimed in claim 17 wherein the first calculation means equalizes, in absolute value, a primary error over two or three pairs of the first pairs of values, and tests that secondary errors for the remaining pairs of values are less than or equal to the primary error, the primary and secondary errors being measured between each pair of values and the linear regression function.

19. Apparatus as claimed in claim 18, wherein the primary error and the secondary errors are weighted with specific weighting coefficients.

20. Apparatus as claimed in claim 18 wherein the first calculation means splits the pairs of values into a plurality of sub-sets, the first calculation means determining one linear regression function for each sub-set and computing the codes for each sub-set, the apparatus comprising means for selecting the sub-set specific to the request value and for addressing to the second calculation means the codes peculiar to the linear regression function of the sub-set selected.