Title of Invention

APPARATUS FOR DETERMINING LIKELIHOOD OF OCCURRENCE OF A CAUSE OF ONE OR MORE EFFECTS

Abstract

Apparatus and method for determining a likely cause or the likehood of the occurence of a cause of one or more effects, in which training data relating to previously identified relationships between one or more causes and one or more effects is used to learn the cause and effect relationship. A number of primary and secondary reference points are chosen in the input space created by belief values representing the strength of effect. A Lagrange Interpolation polynomial (or other function representing the cause and effect relationship) and a weight value is associated with each of the said reference point. Weight values associated with primary reference points are considered as independent variables (primary weight values) and other weight values, which are associated with secondary reference points (secondary weight values, depend (preferably, but not necessarily, linearly) on one or more primary weight values. Belief value in the occurence of likely causes of one or more given effects can be determined using this method or appratus.

Full Text

This invention relates to an apparatus and method for assisting in diagnosing cause and effect in many different circumstances throughout various different disciplines. In recent times, various disciplines, including the manufacturing industry and at least some branches of the medical field, have come under increasing pressure to increase yield, productivity and profits (or reduce costs or overheads). As such, there is an increasing need to reduce costs and reach the required result as quickly as possible, for example, in the manufacturing industry, the required result may be the manufacture of a batch of products which are all of optimum quality, and in a branch of the medical field, the required result might be the correct diagnosis of a complaint or disease in a patient (i.e. in both cases, to get it "right first time"). In the case of the manufacturing industry, manufactured products are usually tested for quality, and sub-standard units are rejected. When a unit or set of units is rejected, the fault or faults are generally noted and it is highly desirable to establish the reason for the occurrence of the faults so that any problems can be rectified and the chances of manufacturing sub-standard products thereafter can be minimised. Such diagnosis is usually performed by experts in the field, who have acquired a fundamental understanding over the years of cause and effect relationships, which often influences the occurrence of rejected product units. Similarly, a patient will usually visit a general practitioner with one or more symptoms, from a description of which the doctor must try and diagnose the problem, based on current medical knowledge and his experience in the field. Thus, in general, cause and effect relationships are learned by experts in a particular field and applied in the diagnosis of problems when they occur in that field. However, this manual diagnosis procedure is time consuming and, therefore, increases costs and reduces productivity and yield. Further, when an expert leaves a particular place of employment, his expertise is also lost to the employer. Attempts have been made in the past to relate belief values, representing the strength of an effect, to a belief value which quantifies the extent of occurrence of an associated cause, using either multivariate regression analysis methods or neural network related methods. If the belief values in 'p' manifestations associated with a cause 'c' are represented by 'p' variables (as shown in Figure 1), then using the multivariate linear regression analysis method [1] or single layered feed forward neural network method [2], the belief value, which quantifies the extent of occurrence of cause 'c' is given by the following equation: where wJ, j = 0 to p are referred to as either regression coefficients (in a regression analysis context) or weights (in a neural network context). These coefficients, or weights, generally being considered as independent variables, are mostly determined using least square minimisation techniques. This is achieved by comparing the belief value in the cause calculated by equation (1) with a previously known value for the same inputs. Multi-layered feed forward neural network techniques, including radial basis functions [2], and also a range of methods proposed in the family of intrinsically linear, multivariate regression analysis [1], generalise equation (1) in the following way: Belief in cause = w0 + w1z1 + w2z2 + w3z3 +... + wizi +... + wmzm (2) where zi (i = 1 to m) represents a function of ?j (j = 1 to p) Different methods and apparatuses may use different functions (zi) ranging from simple linear polynomials to higher order, non-linear polynomials, logarithmic or exponential functions. Although, these methods have been employed to associate belief values in causes to belief values in effects, the methods have two major limitations: 1. Physical interpretation can not be assigned to either weights wi or functions zi, as given in equation (2), which makes it very difficult to gain any physical insight into the cause and effect relationship. 2. The number and the type of function used to define z. are not unique and are determined by a trial and error method. These limitations constrain the applicability of existing methods to relate belief values, which quantify the occurrence of a cause, with belief values representing strength of effects. The inventors have now devised an arrangement that overcomes the limitations/problems outlined above and seeks to provide a generic tool to relate belief values in causes and effects for use in many different fields, and industries, for diagnosing problems occurring therein. Thus, in accordance with the first aspect of the present invention, there is provided apparatus for determining a likelihood of occurrence of a cause of one or more effects, the apparatus comprising means for receiving and/or accessing training data relating to previously-identified relationships between one or more causes and one or more effect means, for defining one or more functions representative of said relationships, said functions) being in the form of polynomials which define quantified mappings between said one or more causes and one or more effects, each of said polynomials being associated with each of a plurality of respective reference points, at least some of which have a weight value assigned thereto, at least one of said weight values being an independent variable ("primary weight") and at least one of said weight values being dependent on one or more of said primary weights "(secondary weight"), and means for determining a likelihood of occurrence of one or more causes of one or more given effects using said mappings. Also in accordance with the first aspect of the present invention, there is provided a method of determining a likelihood of occurrence of a cause of one or more effects, the method comprising the steps of receiving and/or accessing quantifying data relating to previously-identified relationships between one or more causes and one or more effects, defining one or more functions representative of said relationships, said function(s) being in the form of polynomials which define quantified mappings between said one or more causes and one or more effects, each said polynomials being associated with each of a plurality of respective reference points, at least some of which have a weight value assigned thereto, at least one of said weight values being an independent variable ("primary weight") and at least one of said weight values being dependent on at least one of said primary weights ("secondary weight"), and determining a likelihood of occurrence of one or more causes of one or more given effects using said mappings. Thus, the present invention not only provides a method and apparatus for automatically providing diagnostic information relating to one or more likely causes of one or more given manifestations or effects, thereby reducing the input required by expert personnel once sufficient training data has been entered, where human input can otherwise be time consuming and is, of course, prone to error as it is generally based only on the opinion of very few experienced personnel; but also significantly reduces the number of independent weight values required to facilitate the diagnostic process, thereby reducing the quantity of training data required to a practical level when a large number of effects or manifestations of a cause are involved. The present invention provides a method and apparatus for calculating a belief value which quantifies the extent of occurrence of a cause, given belief values which quantify the occurrence, or non-occurrence, of associated affects of the cause. Examples of effects of a cause, which are indicative but certainly not exhaustive, are "symptoms" shown by patients in the medical domain, "defects" occurring in components in manufacturing industry or "effects" as generally meant in any "cause and effect" diagram. The belief value quantifying the occurrence, or non-occurrence, of an effect associated with a particular cause is preferably normalised between +1 to -1 or 1 to 0 respectively. This belief value may also be interpreted as being the strength of the effect. The belief value, which quantifies the extent of occurrence of the cause under consideration, is preferably also normalised from zero to unity or -1 to +1 to represent non-occurrence and occurrence, respectively. The present invention also allows a meaningful physical interpretation to be assigned to each and every said weight value (primary and secondary weights) in the sense that a weight value at a position described by belief values, which represent the strength of associated manifestation, is nothing but the output value i.e. a belief value representing the extent of occurrence of the cause given the strengths of the associated effects. In a preferred embodiment of the present invention, the apparatus comprises a means of producing a multi-dimensional hyper-surface representing belief values in the occurrence /non-occurrence of a cause. The number of dimensions of the hyper- surface is equal to the number of input nodes representing the effects of a cause. The order of the hyper-surface along each dimension is determined by the order of the polynomial, preferably a Lagrange Interpolation Polynomial, used along the said dimension. The first order (or linear) Lagrange Interpolation Polynomial is defined by two reference points. Second order or quadratic Lagrange Interpolation Polynomials require three reference points. Similarly, an nth ordered Lagrange Interpolation Polynomial will require (n + l) reference points along the given dimension. The apparatus or method of calculating reference points including primary reference points and then associating a weight value along with the Lagrange Interpolation Polynomial at each reference point is described in more detail below. The training data is preferably made up of one or more training files, the or each file comprising an input vector, storing belief values representing strengths of all associated effects, and its corresponding desired output vector, storing belief values, which quantify the extent of occurrence of the corresponding cause. An embodiment of the present invention will now be described by way of example only, followed by sample numerical calculation and with reference to the accompanying drawings, in which: Figure 1A and 1B is a schematic illustration of a manifestation cause relationship; Figure 2 A to 2E are graphical representations of some general one-dimensional cause and effect relationships; Figure 3 is a schematic diagram illustrating a two effect - one cause relationship; and Figure 4 is a schematic diagram illustrating a two dimensional space describing ?1 and ?2 axes. In order to facilitate the description of an exemplary embodiment of the present invention, a number of general one-dimensional cause and effect relationships associating belief values (based on the network illustrated in Figure 1B) will first be illustrated. Figures 2A-2E, show possible examples of the variation in belief values, representing the extent of occurrence of cause (output value), with respect to the belief values, representing the strength of one of the associated effect. Referring to Figure 2A of the drawings, a linear variation in belief values is shown, in which when the belief value representing the strength of effect is at its minimum, the belief value representing the extent of occurrence of the related cause is also at its minimum. As the strength of effect increases, the belief value in the occurrence of cause also linearly increases. Referring to Figure 2B of the drawings, a quadratic variation of cause and effect is shown in which when the belief value representing the strength of the effect is at its minimum, then the belief value in the occurrence of the related cause is also at its minimum. As the strength of the effect starts to increase, the belief value in the occurrence of the corresponding cause also starts to slowly increase. As the strength of the effect increases to about half of its maximum value, so the belief value in the occurrence of the cause suddenly increases and reaches its maximum value when the strength of the effect reaches its maximum value. Referring to Figure 2C of the drawings, a quadratic variation of cause and effect is shown which, when the belief value representing the strength of the effect is at its minimum, then the belief value in the occurrence of the related cause is also at its minimum. As the strength of the effect starts to increase, the belief value in the occurrence of the cause quickly starts to increase. When the strength of the effect is around half of its maximum value, the rate of increase in the belief value of the occurrence of the related cause slows down and reaches its maximum value when the strength of the effect also reaches its maximum value. Referring to Figure 2D of the drawings, a quadratic variation of cause and effect is shown in which when the belief value representing the strength of the effect is at its minimum, then the belief value in the occurrence of the related cause is at its maximum. As the strength of the effect starts to increase, there is a quick reduction in the belief value in the occurrence of the cause. As the strength of the effect increases to about half of its maximum value, the belief value in the occurrence of the cause slowly decreases and reaches its minimum value when the strength of the effect is at its maximum. Referring to Figure 2E of the drawings, a quadratic variation of cause and effect is shown in which when the belief value representing the strength of the effect is at its minimum, then the belief value in the occurrence of the related cause is at its maximum. As the strength of the effect starts increasing, the belief value in the occurrence of the cause slowly starts to decrease. When the strength of the effect reaches around half of its maximum value, the belief value in the occurrence of the cause starts to decrease quickly and reaches its minimum when the strength of the efffect is at its maximum. It is an object of the present invention to provide a diagnostic arrangement, which not only "learns" from examples, but also quantifies the cause and effect relationship, which may be described in this exemplary embodiment of the invention by a decision hyper-surface constructed by combining Lagrange Interpolation Polynomials. Two reference points associated with two Lagrange Interpolation Polynomials and two weight values are required to describe a linear variation in the belief values as shown in Figure 2A. Similarly, three reference points along with three Lagrange Interpolation Polynomials and weight values are required to describe one dimensional quadratic belief variation as shown in Figures 2B-2E. The learning function of the apparatus is equivalent to finding a multi-dimensional hyper-surface, describing the said belief variation, which provides a best fit to the training data. Thus, an exemplary embodiment of the invention provides a method for calculating a belief value, which quantifies the extent of occurrence of a cause, given belief values, which quantify the occurrence or non-occurrence of associated effects of the cause. Examples of effects of a cause are, but not limited to, "symptoms" shown by patients in the medical domain, "defects" occurring in components in manufacturing industry or "effects" as generally meant in any "cause and effect" diagram. The belief value ?, which quantifies the occurrence or non-occurrence of an effect associated with a particular cause is normalised between +1 to -1 respectively. This belief value is also interpreted to represent the strength of the effect. The belief value, which quantifies the extent of occurrence of the cause under consideration is also normalised from zero to unity, to represent non-occurrence and occurrence of the said cause respectively. The relationship between the belief value, representing the strength of the effect, and the belief value, representing the extent of occurrence of the cause, is assumed to be either linear, quadratic, cubic and so on. The order of the relationship (e.g. one for linear, two for quadratic, three for cubic etc.) can either be given or calculated iteratively starting from one. To define an nthorder relationship along one-dimension, (n +1) reference points, equidistant between -1 to +1, are chosen. (If the location of the reference points is not equidistant, these reference points are mapped on to another set of equidistant reference points.) For each reference point 'i', a one dimensional Lagrange Interpolation Polynomial is constructed based on the following formula: where, n: order of the Lagrange Interpolation Polynomial (one for linear, two for quadratic, etc.) k. A reference point at which the one-dimensional Lagrange Interpolation Polynomial is constructed, it ranges from 0 to n. i: Ranges from one to total number of reference points i.e. (n +1). are (n +1) equidistant reference points from with (n +1) corresponding Lagrange Interpolation Polynomials as given by equation (3). The variable ?, which stores the belief value representing the strength of the corresponding effect, ranges from -1 to +1. For a "single effect - cause relationship", the Lagrange Interpolation Polynomial is one-dimensional and the reference points are drawn along this dimension. If the number of associated manifestations for a given cause is 'p', the Lagrange Interpolation Polynomial at a reference point 'i' will be 'p' dimensional and is given by the following equation: nj: Order of one dimensional Lagrange Interpolation Polynomial corresponding to jth dimension that represents the relationship between jth manifestation and the cause under consideration. kj: Reference point along jth dimension, at which the one dimensional Lagrange Interpolation Polynomial is evaluated, (kj ranges from 0 to nj.) reference points along the jth dimension. These are the primary reference points as they lie along the jth dimensional axis. i: Ranges from one to total number of reference points 'q' As kj independently varies from 0 to nj for each Lagrange Interpolation Polynomial, q = (n1+l)*(n2+l)*(n3+1)*...*(nJ + 1)*...*(np+1) (6) Thus, the co-ordinates for a reference point 'i', corresponding to each dimension, are given as (k1,k2,...,kj,...,kp). The reference points, which lie along the dimensional axes, are special points and are referred to as "primary reference points". In terms of co-ordinates, a reference point is along the dimensional axis if one and only one of its co-ordinates (kj) has a non-zero value and all other co-ordinates is zero. A weight variable with values constrained between zero and unity is associated with each reference point. Therefore, the total number of weights is the same as the total number of reference points 'q' A weight value at a reference point, in the context of this invention, is considered to be representative of the belief value in the cause. Weights corresponding to primary reference points are chosen as primary or independent weights and the remaining weights - referred to as secondary weights - are dependent on, and expressed as a linear combination of primary weights. This dramatically reduces the number of unknown variables within the network, thereby reducing the amount of training data required to a practical level as compared with conventional techniques. For example, for a 'p' dimensional problem, the number of During the learning or training process, the optimal values for primary weights (constrained between zero and one) and coefficients used in the linear combination expression are determined based on any user defined method including generally available optimisation principles. The secondary weights are also constrained between zero and one. The belief value in the occurrence of a cause, based on the known belief values quantifying the strength of 'p' associated effects, is given by the following equation: where, q: Total number of reference points. wi.: Weight variable associated with the ith reference point. In an embodiment of the present invention, a secondary weight is defined as a linear combination of primary weights. However, the following particular cases of linear combination may also be used in other preferred embodiments of the present invention. 1. A secondary weight value associated with a reference point is a linear combination of those primary weight values, which are associated with primary reference points corresponding to the co-ordinates of the said reference point. 2. A secondary weight value associated with a reference point is a constant multiplied by the average of primary weight values, which are associated with primary reference points corresponding to the co-ordinates of the said reference point. A Numerical Example of an Exemplary Embodiment of the Present Invention Two effects, ?1 and ?2, are associated with a cause c. Therefore, two-dimensional Lagrange Interpolation Polynomials li(?1,?2) will be used for defining the hypersurface. Quadratic relationship is assumed between belief values for the effect ?1 and the cause c, and also for the effect ?2 and the cause c. For this example, the belief value in the occurrence of cause c is calculated for a belief value in the first effect ?1 equal to 0.5 and a belief value in the second effect ?2 equal to-0.5. Numbers 1 to 9 in Figure 4 denote equidistant reference points. As a result of quadratic relationship, (?1 and ?2 equal to 2) three equidistant reference points are used along each dimension. Using equation (6), it can be seen that the total number of reference points is 9. Using equation (7), it can be seen that total number of primary reference points is 5. These points are also indicated in the following table (Table 1), which shows the coordinates of all nine reference points in various forms. Weights associated with primary reference points 1, 2, 3, 4 and 7 are primary weights and are also independent parameters. The secondary weight values at locations 5. 6. S and 9 are expressed as linear combination of primary weights and in particular during a learning process are w1, w2, w3, w4, w7 and constant C. The belief value, in the extent of the occurrence of cause, for a given belief value ?1 equal to 0.5 (representing the strength of effect ?1) and a given belief value ?2 equal to -0.5 (representing the strength of effect ?2) is calculated as: Using equations (4) and (5), Lagrange Interpolation Polynomials are constructed and then evaluated at (0.5, -0.5) at all reference points. Lagrange Interpolation Polynomial for reference point 1: If wl =0.0084, w2 =0.1972, w3 =0.4179, w4 =0.1924, w7= 0.7359 and C = 1.5656, then using equations (9), (10), (11) and (12) ws = 0.3050, w6 = 0.4778, wg = 0.7304 and w9 = 0.9032. During a learning or training phase, the predicted belief value (0.3024) is compared with a known belief value in the training data file to calculate the error. Any of the known optimisation method is used to calculate new values of primary weights and coefficients used in the linear combination equations. The new belief value in the cause is again calculated and the process is repeated until a user defined criterion of error minimisation is achieved. The optimal values of primary weights and coefficients in the linear combination equations are stored and used for future applications. Embodiments of the present invention has been described above by way of example only and it will be appreciated by a person skilled in the art that modifications and variations can be made to the described embodiments without departing from the scope of the invention. WE CLAIM: 1. Apparatus for determining a likelihood of occurrence of a cause of one or more effects, the apparatus comprising means for receiving and/or accessing training data in the form of discrete values relating to previously-identified relationships between N effects (where N=l) and a cause, and means for weight learning, wherein the number of effects N defines the number of dimensional axes of an N dimensional input space and an N-dimensional function lt that defines a decision hyper-surface representative of said training data that quantifies the mapping between said cause and said N effects, and being arranged and configured to : select a number of reference points within the said N dimensional input space, a predetermined number of said reference points being designated as primary reference points and the rest of said reference points being designated as secondary reference points, said predetermined number being dependent on the number of effects N, wherein each reference point has a weight value w, assigned thereto, a weight value being representative of a belief value in said cause given an effect, the weight values assigned to said primary reference points being independent variables ("primary weight") and the weight values assigned to said secondary reference points ("secondary weight"') being dependent on one or more of said primary weights; the apparatus comprising : a network comprising one or more input nodes arranged to receive respective input signals, the or each input signal having an input value that quantifies the occurrence or non-occurrence of an effect associated with said cause, and one or more output nodes for generating an output signal representative of a likelihood of a cause of one or more given effects by calculating the result of the said N-dimensional function l, at each reference point within the said N dimensional input space using said input values, said output signal being the combination of one or more primary and secondary weights at the corresponding said reference points associated with one or more said input values. 2. Apparatus as claimed in claim 1, wherein equidistant reference points are selected along each dimension corresponding to each single effect - cause relationship. 3. Apparatus as claimed in claim 1, where non-equidistant reference points are selected along one or more of said dimensions, and wherein said non-equidistant reference points are mapped onto a set of equidistant reference points. 4. Apparatus as claimed in one of the preceding claims, wherein primary weight values are assigned to reference points on dimensional axes of said function(s). 5. Apparatus as claimed in claim 4, wherein secondary weight values are assigned to reference points in the input space defined by the said dimensional axes. 6. Apparatus as claimed in any one of the preceding claims, wherein a p- dimensional Lagrange Interpolation polynomial is constructed in respect of each reference point, said Lagrange Interpolation polynomials defining said quantified mapping said course and said one or more effects, wherein p is the number of effects associated with said cause. 7. Apparatus as claimed in claim 6, further comprising means for combining said Lagrange Interpolation polynomials to construct a multi- dimensional hyper-surface that is representative of said decision hyper- surface. 8. Apparatus as claimed in any of preceding claims in which each secondary weight value is expressed as a linear combination of one or more primary weights. 9. Apparatus as claimed in claim 8, wherein a secondary weight value associated with a reference point is a linear combination of those primary weight values, which are associated with primary reference points corresponding to the coordinates of the said reference point. 10. Apparatus as claimed in claim 8, wherein a secondary weight value associated with a reference point is a constant multiplied by the average of primary weight values, which are associated with primary reference points corresponding to the co-ordinates of the said reference point. 11. Apparatus as claimed in any one of the preceding claims, in which the likelihood of occurrence of one or more causes and/or strength values of one or more effects are sealed between -1 to +1. 12. Apparatus as claimed in any one of the preceding claims, wherein the weight values (primary and secondary) are constrained between 0 and 1. 13. Apparatus as claimed in any one of the preceding claims, wherein the training data is made up of one or more training files, the or each file comprising an input vector storing belief values representing strengths of all associated effects and its corresponding desired output vector storing belief values which quantify the extent of occurrence of the corresponding cause. 14. Apparatus as claimed in any. one of the preceding claims, wherein optimal values of primary weights and coefficients used in linear combination are determined, during a training or learning process, using an optimization technique involving minimising the error in he predicted belief value and the corresponding known belief value as given in the training data. 15. Apparatus as claimed in any one of the preceding claims, wherein a hyper-surface defined by Lagrange Interpolation polynomials is representative of the belief values in the occurrence non occurrence of cause. Apparatus and method for determining a likely cause or the likehood of the occurence of a cause of one or more effects, in which training data relating to previously identified relationships between one or more causes and one or more effects is used to learn the cause and effect relationship. A number of primary and secondary reference points are chosen in the input space created by belief values representing the strength of effect. A Lagrange Interpolation polynomial (or other function representing the cause and effect relationship) and a weight value is associated with each of the said reference point. Weight values associated with primary reference points are considered as independent variables (primary weight values) and other weight values, which are associated with secondary reference points (secondary weight values, depend (preferably, but not necessarily, linearly) on one or more primary weight values. Belief value in the occurence of likely causes of one or more given effects can be determined using this method or appratus.

Documents:

321-CAL-1999-FORM 27.pdf

321-kolnp-2004-granted-abstract.pdf

321-kolnp-2004-granted-claims.pdf

321-kolnp-2004-granted-correspondence.pdf

321-kolnp-2004-granted-description (complete).pdf

321-kolnp-2004-granted-drawings.pdf

321-kolnp-2004-granted-examination report.pdf

321-kolnp-2004-granted-form 1.pdf

321-kolnp-2004-granted-form 18.pdf

321-kolnp-2004-granted-form 2.pdf

321-kolnp-2004-granted-form 26.pdf

321-kolnp-2004-granted-form 3.pdf

321-kolnp-2004-granted-form 5.pdf

321-kolnp-2004-granted-reply to examination report.pdf

321-kolnp-2004-granted-specification.pdf