JPH06175994A - Sorting device using probability logic in disjunctive normal form - Google Patents

Abstract

PURPOSE:To support diagnosis in medical care and a medical engineering field and treatment and control based on it by quantitatively holding sorting rules as rule expression represented by probability logic in a disjunctive normal form which is easy to be understood by people and semiautomatically performing a sorting processing based on them. CONSTITUTION:The sorting rules 3000 described in the disjunctive normal form are set in the rule setting part 1008 of a knowledge conversion part 1006. A rule analysis part 1007 defines the index numbers phi and sigma of a probability logic formula in the disjunctive normal form '1' or '0' depending on whether or not a (k)-th remark is included in a (j)-th disjunction or whether or not the (j)-th disjunction is included in an output item. A parameter 2002 obtained in such a manner is set at a parameter part 1002. A probability conversion part 1005 holds the relation between an input probability and a measured value which is a constant value and when the measured value 2000 is obtained, sets a corresponding probability value in the input part 1001 of a sorting part 1000. An output probability calculation part 1003 calculates an output probability 2003 from the parameter 2002 and the input probability 2001.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a classification device using disjunctive standard form stochastic logic, and in particular, classification rules including exception rules such as individual differences (or diagnostic rules, judgment rules) are used as probabilistic logic easy for humans to understand. By holding and automatically classifying (or diagnosing, making decisions) using this, and showing the target value of the input in the reverse direction from the output you want to obtain, diagnosis in the medical and medical engineering fields, and based on it The present invention relates to a device that supports treatment / control.

[0002]

2. Description of the Related Art A general-purpose expert shell is a device for describing and using classification rules as rules that are easy for humans to understand. In this method, the knowledge of an expert is described as a rule, and the classification process is performed by sequentially performing case classification using the rule. The general-purpose expert shell is described in detail by Fumio Mizoguchi and others under the title "Expert System" on pages 139 to 183 of the book published by Fuji Techno System in "AI General Guide", 1987.

Fuzzy logic is an expression for describing and using a quantitative classification rule. This can be done by subjectively determining membership functions that relate concepts such as "high" and "low" to real numbers such as temperature, and using fuzzy inference rules for those membership functions. . Regarding fuzzy logic, he is familiar with the book "Fuzzy theory and its applications" published by Science Co. in 1988 (Masaharu Mizumoto).

Fuzzy neural networks are known as an attempt to semi-automatically learn membership functions and inference rules of fuzzy logic. This is to represent the membership function and the inference rule candidate by a product-sum type neural network, and try to optimize them by using the learning method of the neural network. As an example of a fuzzy neural network, in 1991, "A Study on Fuzzy Model" was used.
ing Using Fuzzy Neural Ne
networks. "A study on fuzzy modeling using neural networks", "Proceedings of the IFE"
There is a fuzzy network announced by Mr. Hirokawa, Mr. Furuhashi, Mr. Uchikawa, and Mr. Tagawa on pages 562-573 of Volume 1 of "S'91""Proceedings of the Eyes 91".

Another device that describes the classification rules as rules that are easy for humans to understand is "The thirth" in 1990.
d annual workshop on comp
operational learning theor
y "" The Third Annual Workshop on
Computational Learning Theory, Proceedings 67-81, Kenji Yamanishi, "A lea
running criterion for stoch
astic rules. There is a list of probability decisions published under the title "Learning Criterion Forst Castic Rules." This expresses the classification rule as a conditional probability expressed by a decision list structure. Another device that learns the classification rules semi-automatically is a feedforward neural network model. The feed-forward neural network model was published in an industrial book entitled "PDP Model: Cognitive Science and Search for Neuron Networks" in 1989. (D. E. Rumelhart, "D. Elamerhart" et al. , Shunichi Amari et al.).

[0006]

A device for supporting diagnosis and treatment / control based on it in the fields of medical care and medical engineering is a quantitative method for guiding a disease name or necessary treatment content based on input of clinical test results and the like. The knowledge of a specialized expert must be retained as a classification rule (including a diagnostic rule and a judgment rule) and can be automatically classified using this. To do that, convert existing rules written by experts into classification rules,
A mechanism for holding the device is also required.

Further, in order to carry out more accurate classification, it is necessary to also hold exceptional rules such as individual differences of patients, characteristics of devices, etc., and be able to perform automatic classification using this.

Further, even if there is no existing rule described by an expert, based on the classification cases classified by the expert,
There is a need for a mechanism that can learn the classification rules behind it and use them to classify.

Similarly, when there is no exceptional rule described by an expert, based on the exceptional classification case made by the expert, the exceptional classification rule behind it is learned, and classification is performed using this. A possible mechanism is needed.

In addition, especially in a medical setting, it is necessary for a human to confirm whether the operation of the classifying device is safe or appropriate, and for that purpose, the classification rule held by the device is presented as a rule understandable by a person. A mechanism to do this is necessary.

Further, particularly in the case of control treatments such as treatment and control, there is a certain target result (classification, diagnosis, judgment result) in advance, and conversely from that target, the input target of physiological index or the like. Since a value is often estimated and an effort is made toward this, a mechanism for deriving an input target value from a certain output target value is necessary.

To summarize the above, 1. The classification rules (or diagnostic rules, judgment rules) can be retained as a rule expression that guarantees quantitativeness and is easy for humans to understand, and based on this, the classification process can be performed semi-automatically. It can be used by converting it to the rule expression. 2. Hold exceptional classification rules such as individual differences and classify using them. 3. It is possible to learn the rule expression semi-automatically by using the classification case. 4. Semi-automatically learn exceptional classification rules from exceptional classification cases, 5. The classification rule held by the device can be extracted as a rule that can be understood by people, and There is a need for a classifier that can obtain the target value of the input in the opposite direction from the desired output, but there is no example that has realized such a device. The knowledge representation of the general-purpose expert shell is basically qualitative and is not suitable for quantitative reasoning. In addition, knowledge acquisition in a general-purpose expert shell requires an interview by a knowledge engineer, classification and arrangement of rules, and description of rules by a program, and lacks a semi-automatic knowledge acquisition function. Fuzzy logic can construct quantitative classification rules that are easy for humans to understand, but the basis for determining the membership function is unclear, and there remains a problem in the quantitativeness of inference rules such as using the maximum and minimum values. It has also been pointed out that skill in setting the membership function is required.

The fuzzy neural network solves the latter problem by utilizing the learning function of the neural network.
The question of quantitativeness of inference remains.

The feedforward type neural network can semi-automatically learn the classification rule by preparing a required number of neuron elements and presenting a learning data set consisting of input values and classification results, which saves labor. , People cannot understand the rules obtained as a result of learning. Therefore, it is difficult for human beings to examine the learning result, it is difficult to guarantee safety, and it is difficult to identify the cause when inconvenience occurs.

As a neural network capable of learning exceptional rules, there is a "write-once learning pattern recognition device" of Japanese Patent Application No. 2-198929, but since the element used for write-once learning is a linear sum type neuron, After all it is difficult to follow the logic.

In addition, in order for the neural network to acquire the classification rule, it is considered through learning from the classification case, and existing knowledge cannot be directly used like the expert system.

The probability determination list is easy for humans to understand and the quantitativeness of inference is guaranteed, but a problem remains in learning ability because a means for efficiently constructing the structure of the list is undeveloped during learning.

Further, the above-mentioned conventional technique lacks means for indicating the target value of the input in the opposite direction from the desired output, and cannot support the actions such as treatment, control, adjustment, etc., in which the target value is provided. That is, the conventional classification system cannot satisfy the above-mentioned six conditions, and cannot support diagnosis in the fields of medicine and medical engineering and treatment / control based thereon.

The objects of the present invention are: The classification rules (or diagnostic rules, judgment rules) can be retained as a rule expression that guarantees quantitativeness and is easy for humans to understand, and based on this, the classification process can be performed semi-automatically. It can be used by converting it to the rule expression. 2. Hold exceptional classification rules such as individual differences and classify using them. 3. It is possible to learn the rule expression semi-automatically by using the classification case. 4. Semi-automatically learn exceptional classification rules from exceptional classification cases, 5. The classification rule held by the device can be extracted as a rule that can be understood by people, and An object of the present invention is to provide a classifier that can obtain the target value of input from the target output in the opposite direction.

[0020]

A classifying apparatus according to the first invention of the present application comprises an input section for inputting an input probability, and a parameter section for holding a parameter of a rule expressed by a disjunctive standard form probability logic. , A rule setting unit that has an output probability calculation unit that calculates an output probability from the parameter and an input probability and an output unit that outputs the output probability as it is or after processing it, and sets an existing classification rule. Section and a rule analysis section that analyzes a set rule to obtain a disjunctive standard form probability logic parameter and sets it in the parameter section. It has a probability conversion part which converts and inputs into the classification part.

According to a second invention of the present application, the classification device of the first invention further holds an exception parameter section for holding a parameter of an exception rule which takes precedence over a given classification rule, and when an input is obtained, An exception determination unit that determines whether or not it satisfies the antecedent part of the exception rule, and if it does, the output probability is modified from the input probability and the parameter of the exception rule and passed to the output unit, and if not satisfied. Has an exception correction unit that passes the output probability to the output unit without correcting the output probability.

In a third invention of the present application, the classification device of the first or second invention further uses the input probability in the classification case consisting of the input probability and the classification result probability, and the output probability output by the classification unit. And an error calculation unit that obtains an error from the classification result probability in the classification data, a parameter correction unit that corrects the parameter of the parameter unit based on the error calculated by the error calculation unit, and an error. Is provided with a rule learning unit including a convergence determination unit that repeats parameter correction until it becomes less than or equal to a preset allowable error.

In a fourth invention of the present application, the classification device of the second invention further uses the exception input probability in the exception classification case consisting of the exception input probability and the exception classification result probability to output from the exception correction unit. An exception error calculation unit that calculates the error between the probability and the exception classification result probability, and an exception that modifies the parameters of the exception parameter unit in the direction of reducing the exception error based on the exception error calculated by the exception error calculation unit The present invention is characterized by having an exception rule learning unit including a parameter correction unit and an exception convergence determination unit that repeats exception parameter correction until the exception error becomes less than or equal to a preset allowable error.

The fifth invention of the present application is the above-mentioned first or second invention.
The classification device of the invention further comprises a rule binarization unit for binarizing the parameter part of the classification part or the parameter of the exception parameter part, and a disjunctive normal form type classification from the binarized parameter. The present invention is characterized by having a rule extraction unit including a rule output unit that outputs a rule.

A sixth invention of the present invention is that the classification device of the above-mentioned first invention further uses the classification unit to inversely estimate an input probability necessary for obtaining a target output probability.
A target output setting unit that sets a target output probability, an input holding unit that holds the input probability, and an error between the output probability output by the classification device and the target output probability using the input probability of the input holding unit ( Recollection error), an input correction unit that corrects the input probability held by the input holding unit in the direction of reducing the recollection error, and an input until the recollection error becomes equal to or less than a preset allowable recollection error. And a recall probability output unit that outputs the input probability held by the input holding unit after the recall error becomes equal to or less than the allowable recall error.

[0026]

The classifying device of the present invention has an input section for inputting an input probability, a parameter section for holding a parameter of a rule expressed by a disjunctive normal form probability logic, and an output probability from the parameter and the input probability. The output probability calculation unit for calculating the output probability and the classification unit including the output unit for outputting the output probability as it is or after processing is output, and the rule setting unit for setting the existing classification rule and the set rule are analyzed. Probability of obtaining a parameter of a disjunctive normal form type probabilistic logic, converting the measured value into an input probability, and inputting to the classification unit, having a knowledge conversion unit consisting of a rule analysis unit set in the parameter unit. Since the conversion unit is provided, the classification rules (or diagnostic rules, judgment rules) can be retained as a rule expression that guarantees quantitativeness and is easy for humans to perform semi-automatic classification processing based on this. , Existing Classification rules, can be used to convert to the rule representation.

Furthermore, an exception parameter part that holds parameters of an exception rule that takes precedence over a given classification rule, and an exception that, when input is obtained, determines whether or not it satisfies the antecedent part of the exception rule. An exception that is passed to the output unit after modifying the output probability from the judging unit and the input probability and the parameter of the exception rule if the condition is satisfied, and passes to the output unit without modifying the output probability if not satisfied. Since it has a correction part,
Exceptional classification rules such as individual differences can be retained and used for classification.

Further, an error calculation unit for obtaining an error between the output probability output by the classification unit and the classification result probability in the classification data using the input probability in the classification case consisting of the input probability and the classification result probability, and the error Based on the error calculated by the calculation unit, a parameter correction unit that corrects the parameter of the parameter unit in the direction of decreasing the error, and a convergence determination that repeats the parameter correction until the error becomes equal to or less than a preset allowable error. Since it has a rule learning section consisting of a section and a section, the rule expression can be learned semi-automatically by using a classification case.

Furthermore, the error between the output probability output by the classifier of the third invention and the exception classification result probability is calculated using the exception input probability in the exception classification case consisting of the exception input probability and the exception classification result probability. An exception error calculation unit, an exception parameter correction unit that modifies the parameters of the exception parameter unit based on the exception error calculated by the exception error calculation unit to reduce the exception error, and an exception error preset Since it has an exception rule learning unit including an exception convergence determination unit that repeats the correction of the exception parameter until the error becomes less than or equal to the error, the exceptional classification rule can be learned semi-automatically from the exceptional classification case.

Further, a rule binarization unit for binarizing the parameter part of the classification part or the parameter of the exception parameter part, and a disjunctive normal form type classification rule are output from the binarized parameter. Since it has the rule extraction unit including the rule output unit, the classification rule held by the device can be extracted as a rule understandable by a person.

Further, a target output setting unit for setting a target output probability in order to inversely estimate an input probability necessary for obtaining a target output probability by using the classification unit,
An input holding unit that holds the input probability, a recall error calculation unit that obtains an error (recall error) between the output probability output by the classifier and the target output probability using the input probability of the input holding unit, and the recall error The input correction unit that corrects the input probability held by the input holding unit in the direction of decreasing, the recall convergence determination unit that repeats the correction of the input until the recall error is equal to or less than the preset allowable recall error, and the recall error is Since it has the recall probability output unit that outputs the input probability held by the input holding unit after the error falls below the allowable recall error, it is possible to reverse the target value of the input from the target output.

[0032]

BEST MODE FOR CARRYING OUT THE INVENTION The classification device of the present invention supports diagnosis and treatment / control based on it in the fields of medical treatment and medical engineering. First, examples of the support will be shown, it will be shown that they can be realized by the classification device of the present invention, and then concrete implementation means of each invention will be described. A blood sugar level control system for diabetic patients will be taken up and explained as an example of means for supporting diagnosis and treatment in medical care.

This system is a device for always maintaining a normal blood glucose level by wearing it on a diabetic patient or implanting it in the body, but this system has not been realized yet, but it is strongly desired to be realized. System. Some diabetes
Some of these are caused by a decrease in the secretion of insulin, a hormone that lowers the blood sugar level. Diabetes mellitus in this case involves the administration of insulin into the blood so that the blood sugar level is kept in the normal range (about 100 mg / dl). It is known that the symptoms can be reduced. (Showa 56,
Under the title of Medical Physiology Perspective, W. F. Two of the books published by Maruzen Publishing Co., Ltd., written by Ganong, W, F, Gannon, translated by Kojiro Matsuda and others.
(See pages 87-299 for details).

However, the normal blood glucose level of a patient is not always constant, but changes greatly with meals, cold stimuli, exercise, and the like.
In addition, since the strength (sensitivity) of a certain amount of insulin varies from individual to individual, diagnosis and treatment cannot be successfully performed simply by measuring the blood glucose level.

The patient's condition, for example, whether he / she is exercising or after eating, is multilaterally diagnosed by measuring other physiological visual targets such as body temperature and digestive hormone concentration. And the individual's susceptibility to insulin needs to be considered as an exceptional rule.

Given the test values such as blood glucose level, body temperature, digestive hormone, and insulin concentration, the specialist can determine the amount of insulin to be administered next based on the test values, but the specialist always accompanies the diabetic patient. Since this is impossible, a device that can substitute the diagnosis / treatment knowledge is required.

According to the first aspect of the present invention, the rules described for determining the condition of the patient, determining the dose of insulin, etc., for example, "If the patient is kept in a cold state, the blood sugar level is slightly increased. It is possible to control the insulin administration device using the diagnostic result based on the conversion of a rule such as “doing a small dose of insulin” into a quantitative rule and holding it.

As an administration device, a small tank of insulin is used to administer into a blood from a catheter placed in a patient's vein through a silicon tube and a micropump,
As for the dose, a device is generally used in which the rotation speed of a micro pump is controlled. For example, by setting the rotation speed of the pump for each diagnostic result,
Diagnosis and treatment are realized.

Further, when the second invention is used, the device can maintain an insulin sensitivity of an individual patient, for example, an exceptional rule such as "this patient is fully effective with half the usual amount of insulin". Based on this, it becomes possible to control the dose of insulin to be reduced by half.

Further, for example, when a doctor looks at a test value, he thinks a great many things in an instant, but it is often difficult for the doctor to describe such work.

Even in such a case, if the third and fourth inventions are used, the diagnosis and treatment rules of the doctor behind them are learned and held only from the diagnosis and treatment cases, and the diagnosis and treatment are used for diagnosis and treatment. I can help. Further, in order to confirm safety, the diagnosis / treatment rule held by the system needs to be regularly checked by a doctor, but according to the fifth invention, the held rule is presented to the doctor in an interpretable description. You can do the final safety confirmation. In addition, when the amount of insulin in the insulin tank is very low, or when the remaining battery power of the pump is low, the amount of insulin that can be administered is determined on the contrary. On the contrary, how should the patient behave? For example, “Because the possible dose of insulin decreased per minute,
Patients should be advised to keep the temperature below cal / min and avoid cold conditions and limit sweet foods to 20 grams. "

This function can be realized by the sixth invention. INDUSTRIAL APPLICABILITY The present invention can be applied not only to the example of the blood sugar level control system for diabetes described above, but also to the control of the pacemaker of a patient with heart disease, the rapid diagnosis at the time of mass examination, and life guidance. In such a case, a function equivalent to that of the blood glucose level control system described in the example is required, except that the physiological optotype, the diagnostic item, and the control target to be measured are different, and those functions can be realized by the first to sixth inventions. .

Next, a concrete realizing means of the classification device of the present invention will be described.

First, the "disjunctive standard form probabilistic logic", which is an expression of a classification rule (diagnosis rule, judgment rule) used by the classification device of the present invention, will be briefly described. The sample space Ω is a set of specialists (doctors) ω of a limited number (M persons), and the root probabilities that represent the correctness of the diagnosis to each specialist

[0045]

[Equation 1]

Allocate (Note: P (ω) = 1 / M is assumed because the superiority or inferiority of the expert cannot be determined at first, but if the correctness of the diagnosis can be determined by following up the diagnosis case, the diagnosis result will be based on this. The p of the expert who was right
(Ω) is higher and the p (ω) of the expert who was wrong is changed to a lower one, so that the diagnostic result obtained from M people is more correct. Findings as a random variable on this space Χi (R, ω) and diagnostic Xo (R, ω) are introduced. X
i is a certainty factor ([0,1]) given to an individual finding item, for example, “having fever”, when an expert views the test value vector R such as blood pressure, body temperature, and blood glucose level, Xo
Is a certainty factor ([0,1]) given to a diagnostic item, for example, "a cold".

When Xk is an individual element of the finding or diagnosis, the event is E = {ω | Tk≤Xk (ω)} (Tk
Is a threshold value), and the probability of a certain event E is

[0048]

[Equation 2]

It is shown by. The diagnostic result we want to obtain is not the subjective value Xo given by one expert, but p (Eo) obtained from Ω. When the event corresponding to the k-th finding item is Eik and the event corresponding to the i-th diagnostic item is Eoi, the process of diagnosis (classification) is R → Xi → p (Ei) →
It can be represented by a process of p (Eo).

R → Xi → p (Ei), Xo → p (Eo)
The relationship can be determined experimentally, and because the names of findings and diagnostic items are clear, interpretation is easy, but p (Ei) → p
Interpretation of (Eo) is generally difficult. To enable its interpretation, we use the quantitative logical expressions discussed below.

The set arithmetic ∪, ∩, − satisfy the formulas of the propositional arithmetic, and all the expressions written by ∪, ∩, − are disjunctive normal form (D
NF: Disjunctive Normal For
m).

If enough observation items are prepared, Eo can be written as Eo = ∪∩Ei in the set calculation of DNF format, and (Eo) = p
(∪ ∩ Ei). by the way

[0053]

[Equation 3]

Therefore, in order to find p (Eo), it is necessary to find the probability (joint probability) of the product set over two steps. However, since it is generally difficult to obtain the joint probability that contributes to diagnosis from the data, the marginal probability is used to approximate the following equation.

[0055]

[Equation 4]

If the logarithm of both sides of (1) is taken,

[0057]

[Equation 5]

Λk is a weight for expressing the log-likelihood of the joint probability using that of the marginal probability, that is, the “magnitude of contribution to the joint probability” of each marginal probability. Eik having a large λk influences ∩Ei more than others. Conversely, λk
Eik that is close to 0 does not influence ∩Ei. In other words, it can be said that it is not included in the product event. Using the method of (1),

[0059]

[Equation 6]

(2) is referred to as disjunctive normal form stochastic logic (or extended DNF).

This equation can be easily multiplexed

[0062]

[Equation 7]

Σ represents which product ∩ contributes to the sum ∪,
φ represents which marginal probability contributes to the product ∩. By holding the parameters σ and φ, the output probability p (Eoi) can be obtained based on the input probability p (Eik).

The following examples are the ones or a plurality of products manufactured by NEC Corporation, which have software stored thereon.
A workstation equivalent to the workstation EWS-4800 series, or a computer 001 which is a personal computer equivalent to the personal computer PC-H98 model 100, a keyboard attached to these, an RS-232C board to be additionally used, a GPIB board, an Ethernet board , Input means 00 such as NEC-PC-H98-B04
2. Display attached to computer 001, RS-232C board used additionally, GPIB board, Ethernet board, for example PC-H98 of NEC Corporation
-It can be realized by the output means 003 such as B04. FIG. 1 is a basic configuration diagram of the first invention. The classification device of the first invention comprises a knowledge conversion unit 1006, a classification unit 1000, and a probability conversion unit 1005.

The classification unit 1000 has an input unit 1001 for inputting an input probability 2001, and σ and φ of the disjunctive normal form probability logic.
Parameter section 100 holding a parameter 2002 consisting of
2. An output probability calculation unit 1003 that calculates an output probability 2003 from the parameter 2002 and the input probability 2001,
An output unit 1004 that outputs the output probability 2003 as it is or after processing it to output it as a classification result 2004.

The knowledge conversion unit 1006 converts the described existing rule into a parameter 2002 to convert it to the parameter unit 1002.
The rule setting unit 1008 that sets the existing classification rule 3000 and the rule analysis unit 1007 that analyzes the classification rule 3000 to obtain the parameter 2002 and sets the parameter 2002 in the parameter unit 1002.

The probability conversion unit 1005 converts the measured value 2000, which is a real value, into the input probability 2001, and the input unit 100
Enter 1.

First, the rule setting unit 1 of the knowledge conversion unit 1006
In 008, the classification rule 3000 described in the disjunctive normal form is set. The classification rule described in the disjunctive normal form means that a finding item such as “a blood sugar level is slightly high” is Ak (k = 1,
When expressed as 2 ...), "(A1 and A2 and A3)
Or (A4 and A5) or. ． If it is B, then the word connected by "and" is "or".
Refers to the rule that has the antecedent expression added together in.

The rule analysis unit 1007 first determines the classification rule 30.
The total number of finding items in the antecedent part of 00 (the total number is k) and the number of disjunctions (the total number is j) are examined. Next, it is examined which finding item is included in each disjunction, and φjk is set to 1 when the j-th disjunction includes the k-th finding item, and is set to 0 when it is not included. When each output item includes the j-th disjunction, σij is set to 1, and otherwise, it is set to 0.

The parameter 2002 (σ,
(set of φ) is set in the parameter section 1002. The setting of the classification rule 3000 in the rule setting unit 1008 is performed by the input means 00.
2 and the processing of the rule analysis unit 1007 can be realized by the computer 001. The probability conversion unit 1005 holds the relationship between the measured value that is a real value and the input probability, and when the measured value 2000 is obtained, sets the probability value corresponding to this to the input unit 1001 of the classification unit. The input of the measured value 2000 can be realized by the input means 002, and the conversion processing into the input probability 2001 can be realized by using the same computer 001 as that which configures the rule analysis unit.

Output probability calculator 1003 of classifier 1000
Is the parameter 2002 set in the parameter section 1002.
And from the input probability 2001 set in the input unit 1001,
The output probability 2003 (corresponding to p (Eoi)) is calculated based on the equation (3) and is sent to the output unit 1004. Output unit 100
4 outputs the output probability 2003 as graphics, or holds the relationship between the output probability 2003 and the control amount, and outputs the classification result 2004 as the control amount corresponding to the output probability 2003. Input unit 1001, parameter unit 10
02, the output probability calculation unit 1003, the rule analysis unit 100
7. The same computer 001 that realized the probability conversion unit 1005 is used, and the output unit 1004 is output means 00.
It can be realized by using 3. Next, the means for realizing the second invention will be described, but before that, the expression of the exceptional rule handled by the second invention will be described.

Here, the exceptional rule is expressed by an additional learning element (ALU: Additive) by an AND type unit.
Learning Units) is used.

The basic structure of the AND type unit is

[0074]

[Equation 8]

It is mathematically the same as the expression showing the contribution of each finding item to the disjunction in the disjunctive standard form probabilistic logic.

From equation (4),

[0077]

[Equation 9]

The output of the AND type unit is proportional to the inner product of the input weight vector and the vector L of log (Eik), k = 1, 2, ..., K. Since log (Eik) is all 0 or less and 0 ≦ wk, wk = −log p (Eik) / | L | (5) and consider a unit ALU that outputs Q = 1−q.

As a result, q becomes minimum when the vector of ω and L are antiparallel, and then the output Q of the ALU becomes maximum. When certain exception data is obtained, ω is determined according to (5), and if this is held, L is stored. Q
Modifies the output probability 2003 via the output connection weight γib (b: the number attached to the ALU).

Excitatory ALU and inhibitory ALU are distinguished according to the sign of γib.

The rule of the classification unit including ALU is

[0082]

[Equation 10]

Here, f [x, θ] is a threshold function newly defined for the ALU. X is output only when x exceeds the threshold θ (called firing), and 1 is output otherwise. Is a function. By using this threshold function, the ALU also reacts to data that is close to previously learned individual difference data.

In such a case, that is, the input probability 2001
If f satisfies the antecedent part of the exception rule, f [] will fire.

FIG. 2 is a basic block diagram of the classification device of the second invention. The shaded portion is different from the first invention.

The second invention has an exception parameter section 1010, an exception judgment section 1011 and an exception correction section 1009 in addition to the classification section 1000.

In the exception parameter section 1010, the exception rule parameter 2005 (ω, γ) is first input using the input means 002.
To set. The exception determination unit 1011 is a part that performs the above threshold function of f [], and if the qb calculated from the exception parameter 2005 and the input probability 2001 according to the equation (7) is larger than θb, the exception signal 2007 composed of qb.
Is passed to the exception correction unit 1009. The exception correction unit 1009
When the exception signal 2007 is received, the output probability 2003 is corrected according to the equation (6), and then the corrected output probability 3002 is passed to the output unit 1004.

If the exception signal 2007 is not received, no correction is performed, and the output probability calculated according to the equation (3) is passed as it is to the output unit 1004 as the corrected output probability 3002. An exception parameter section 1010 and an exception determination section 1011
The exception correction unit 1009 can be implemented using the same computer 001 that implements the classification unit 1000.

Next, the classifying device of the third invention will be described, but before that, the learning system will be described. Learning disjunctive normal form stochastic logic is the error calculated using expert classification cases.

[0090]

[Equation 11]

Is to search and determine a set of φ and σ that minimizes under the constraint of 0 ≦ φ and 0 ≦ σ.

For this, the well-known non-linear optimization method such as the steepest descent method, the Powell method, or the conjugate gradient method can be used. The above constraint can be satisfied by using the interior point penalty method.

Steepest descent method, conjugate gradient method, Powell
The method and interior point penalty method are
Pp. 140-146, 18 of the book published by Nikkan Giren in 1979 by Hiro Konno and Hiroshi Yamashita, respectively.
Pages 5 to 192, 175 to 185, 220 to 22
See page 4 for details.

As an example, a method combining the steepest descent method and the interior point penalty method will be described. The steepest descent method is a method of observing the gradient of the evaluation function that is minimized and correcting the parameter in the direction in which the gradient steeply descends. The gradient is evaluated by the first-order partial differential coefficient obtained by differentiating the evaluation function with the parameter to be optimized.

In the interior point penalty method, a penalty term is added to the evaluation function to construct a new evaluation function that becomes infinite when it is likely to deviate from the constraint condition, and optimization is performed on that function. Is. Usually, a small coefficient τ (level sequence) is applied to the penalty term, and when it approaches the minimum point, τ is gradually decreased to finally reduce the minimum point of the original evaluation function without the penalty term. Is required.

The number of finding items is k, the number of disjunctions is J, the number of classification items is |, the number of classification cases is N, and the input probability data is x (k
n), k = 1, 2, ..., N classification result data is t (i
n), i = 1, 2, ..., I, output probability 2003 is s (i
n), the output squared error is

[0097]

[Equation 12]

The calculation formulas for the steepest descent of this evaluation function are shown below.

[0099]

[Equation 13]

The parameter correction amount Δ (t) is

[0101]

[Equation 14]

There is also a method for starting the search for φ and σ by selecting an appropriate J and setting φ and σ using random numbers. In addition, in the third invention, a new classification example is provided. From
A method of estimating initial values of J and φ and σ is used. This method will be described below. When Eik, -Eik and their intersection are represented by F, each element of F can be represented by one disjunction. Since Eoi is represented by the sum of the elements of F, the number J of disjunctions is the number of F included in Eoi among individual data.

Since one disjunction can represent at least one classification case, the required J is smaller than the number N of classification cases. Conversely, optimistically, an expert who diagnoses with simpler rules gives the correct diagnosis result: (A∩B) ∪ (A∩-B)
= A, (A ∩ B) ∪ A = A Considering J (inf) obtained by reducing disjunctions as much as possible, the actually required J will be somewhat larger than that. So J
Can be estimated to be in the range of J (inf) ≦ J ≦ N.
Hereinafter, a method for obtaining J (inf) will be described.

Each element of F includes 1: Eik, 0:-
C = I (1) I (2) I (3). ． I (k), I
It can be expressed by a code of (k) = 1 or-1.

Now, two elements B1 and B2 are taken out from F, and only one character of the codes C1 and C2 expressing each is such that one is 1 and the other is 0, or at least one is -1, and the other character strings are Let's think that all are equal. At this time, B1∪B2 = D (D is the intersection of B1 and B2), and B1 and B sum can be expressed only by the code representing D. Using this, an algorithm for estimating J (inf) is configured below.

J (inf) estimation algorithm Aj 1: J is set to 0. 2: Code new data, obtain Cnew, and go to 3.
Stop when all data is over. 3: Among the held Cols, d = | Cold−Cnew | is obtained in order from the one that is not compared with Cnew, and the process proceeds to 4. Go to 5 after comparing all the Colds. 4: ・ If d = 0, go to 2.・ If d = 1, the different character is set to -1, and the code is retained. Delete the Cold just compared and go to 2.・ If d> 1 and one of the different characters is -1, go to 2. Otherwise go to 3. 5: Add 1 to J. Hold Cnew. Go to 2. When Aj is stopped, J encodes the number of disjunctions required, and the individual Colds held hold the product that the disjunction represents. If Aj is used, the maximum number is n (n-1) / 2, where N is the number of data.
The number of disjunctions J can be determined by comparing the number of times.

The initial values of φ and σ can also be obtained by using Cold. Assuming that an event determined by a certain finding and its residual event are evaluated at the same time, the input node is 2k.
1) σi
Let j be all 1. The kth character of the jth Cold (corresponding to the jth disjunction) is

[0108]

[Equation 15]

The following operation may be performed.

FIG. 3 is a basic configuration diagram of the third invention.

The shaded portion is different from the first and second inventions.

The rule learning unit 1012 of the third invention uses the input probability 2001 in the classification case 2010 consisting of the input probability 2001 and the classification result probability 2012, the output probability 2003 output by the classification unit, and the output probability 2003 in the classification data. Error calculation unit 101 for obtaining an error 2008 with the classification result probability 2012
3 and the error 2008 calculated by the error calculation unit, the parameter correction unit 1014 that corrects the parameter 2002 of the parameter unit 1002 in the direction of decreasing the error.
And the convergence determination unit 101 that repeats the correction of the parameter 2002 until the error becomes equal to or less than the preset allowable error.
It consists of 5.

A case of learning using the steepest descent method and the interior point penalty function method will be described as an example. First, classification example 20
10 and the allowable error 3003 to the convergence determination unit 1015, the initial error value 3004 to the error calculation unit 1013, and the input means 00.
Set using 2. Next, the convergence determination unit 1015 finds the search start points of φ and σ using the above-described method using random numbers or the algorithm Aj, and the parameter unit 100
Set to 2. Next, the error calculation unit 1013 sets the set error initial value 3004 as the error 2008 and sets the convergence determination unit 1
Pass it to 015.

The convergence determination unit 1015 detects the error 2008,
The allowable error 3003 is compared, and if the error 2008 is larger than the allowable error 3003, the classification case 2010 to be held is held.
The input probability 2001 is sent to the input unit 1001, and the classification result probability 2012 forming a pair with the input probability 2001 is sent to the error calculation unit 1013 until all classification cases are completed.

After sending, the case presentation end signal 300
5 is sent to the error calculation unit 1013. Error calculator 1013
Is an error 2008 (t (in) -s () based on the equation (10) based on the output probability 2003 calculated by the classification unit 1000 from the input probability 2001 and the classification result probability 2012 sent from the convergence determination unit 1015. in)). In the case of the Powell method without using the differential coefficient, H based on the equation (8) is calculated as the error 2008. ) Is calculated, and when the case presentation end signal 3005 comes, the parameter correction unit 101
Send to 4.

In the parameter correction unit 1014, the error 200
8 and the parameter 2002, the correction amount is calculated based on the equations (9) and (10), and this is added to the parameter 2002 to obtain a new parameter 2002.
It is newly set in the parameter section 1002.

At this time, if the error 2008 is sufficiently small (for example, 1.OE-3 or less), τ is gradually reduced by, for example, reducing it to 1/10. Error calculator 1013
The error 2008 from is sent not only to the parameter correction unit 1014 but also to the convergence determination unit 1015.
5 compares the error 2008 with the allowable error 3003. If the error 2008 is smaller than the allowable error 3003, the correction stop command 2011 is sent to the error calculation unit 1013 and the parameter correction unit 1014 to stop the activity.

Otherwise, the convergence determination unit 1015, the classification unit 1000, the error determination unit 1013, the parameter correction unit 1
014 repeats the above processing. Correction stop instruction 2011
Received error calculation unit 1013, parameter correction unit 1014
Ceases activity.

As described above, the processing of the rule learning unit 1012 for learning the diagnostic rule from the classification case 3003 can be realized.

Setting of classification case 2010, allowable error 300
3 and the initial error value 3004 are set by the input means 002.
Can be realized by using a computer, and each unit of the rule learning unit 1012 can be realized by a computer 001. Next, a method of realizing the fourth invention will be described, but before that, learning of exception rules will be described.

The expert uses (p (Ei
When k) and ti) are given,

[0122]

[Equation 16]

If exceeds the error upper limit ε, additional learning is performed. Although it is conceivable to add the above-mentioned ALU each time the individual difference data is obtained, the following processing Aa is performed in order to increase the number of ALUs as little as possible.

Exception Rule Learning Algorithm Using ALU Aa I. If si << ti: 1: suppressive AL during firing connected to the i-th output node
Raise the U threshold to suppress ignition. When the firing threshold of the ALU is θb, and the output value of the ALU before f [] is acting is q′b, q′b + δ, (0 <δ) is a new θ.
b. δ is a small non-negative real number. If Er (i) <ε, stop. If this is not enough, go to the next step. 2: Excitatory AL during firing connected to the i-th output node
Strengthen the connection weight of U. When the total number is represented by Bn and the other disjunctions and the output of ALU are collectively represented by C,

[0125]

[Equation 17]

Then,

[0127]

[Equation 18]

Assuming that each excitatory ALU is equally responsible, γib = (log (1-ti) -log C) / (Bn
log qb). If Er (i) <ε, stop. If this is not enough, go to the next step. 3: Add excitatory ALU. The input weight ωbk of the ALU to be added is log P (Eik)
And the output weight γib is determined by the same formula as the above-described enhancement of the coupling weight. II. In the case of si >> ti 1: Excitatory AL during firing connected to the i-th output node
Raise the U threshold to suppress ignition. If this is not enough, go to the next step. 2: Inhibitory AL during firing connected to the i-th output node
Strengthen the connection weight of U. As a result, if Er (i) <ε, then stop. If this is not enough, add inhibitory ALU.

FIG. 4 is a basic block diagram of the classification device of the fourth invention.

The shaded portion is different from the first, second and third inventions.

The fourth invention is an exception classification case 201 composed of an exception input probability 2015 and an exception classification result probability 2018.
An exception error calculation unit 1016 for calculating an error between the output probability 2003 output by the classification device of the second invention and the exception classification result probability 2018 using the exception input probability 2015 in 4 above.
And the exceptional error 201 calculated by the exceptional error calculating unit 1016.
9, the exception error correction unit 1017 for correcting the exception parameter 2005 of the exception parameter unit 1010 in the direction of decreasing the exception error 2019, and the exception error 20
An exception convergence determination unit 1018 that repeats the correction of the exception parameter 2005 until 19 becomes equal to or less than a preset exception tolerance 3006. First, exception classification example 20
14 and the exception allowable error 3006 are set to the exception convergence determination unit 1018.
The exception error initial value 3007 is set to the exception error calculation unit 1016.
Set to. Also, all the parameters in the exception parameter section 1010 are deleted.

Next, the exception error calculation unit 1016 sends the exception error initial value 3007 as the exception error 2019 to the exception convergence determination unit 1018.

1) The exception convergence determination unit 1018 compares the exception error 2019 with the exception allowable error 3006, and when the exception error 2019 is larger than the exception allowable error 3006,
One of the exceptional input probabilities 2015 of the exceptional classification case 2014 is sent to the classification device of the second invention as an input probability 2001,
At the same time, the exception classification result 2018 is sent to the exception error calculator.
The exception error calculation unit 1016 obtains the exception error 2019 from the output probability 2003 of the classifying device of the second invention and the exception classification result 2018, and sends it to the exception parameter correction unit 1017 and the exception convergence determination unit 1018. Exception parameter correction unit 10
Reference numeral 17 denotes an exception error 1016, a calculation intermediate result 3008 of the output probability calculation unit 1003 (corresponding to C in the explanation of Aa), the current exception parameter 2006, and the exception parameter 2006 based on Aa. 20
17 is obtained and set in the exception parameter section 1010. The exception parameter modification unit 1017 sends a modification end signal 3009 to the exception convergence determination unit 1018 every time modification is performed.

The exception convergence determination unit 1018 determines the exception error 20
If 19 is larger than the exception allowable error 3006, when the correction end signal 3009 is received, the same exception input probability 2015 as before is sent to the classification device of the second invention as the input probability 2001. If the exception error 2019 is smaller than the exception allowable error 3006, the same operation is performed from 1) for the next exception case data. By repeating this until the exception error becomes less than or equal to the exception allowable error 3006 for all the exception case data, the learning of the exception rule can be realized.

Exception case 2014 setting, exception tolerance 3
The setting of 006 and the exception error initial value 3007 can be realized by the input unit 002, and the exception convergence determination unit 1018,
Exception error calculation unit 1016, exception parameter correction unit 1017
Can be realized by the computer 001.

Next, the means for realizing the fifth aspect of the present invention will be described. Before that, a method of predicating the classification rules held by the classification device of the present invention as rules that are easy for humans to understand will be described. First, a method of predicating the classification rules expressed by the disjunctive normal form probability logic will be described.

All Boolean functions can be converted to DNFs, which can easily be translated into ".. and .. or .. and ..." Predicates. Here, it is considered that the predable person-friendly rule is a Boolean function, and conversion of the extended DNF into a Boolean function is considered.

As a special case of the extended DNF, P (Ei
k) = 0or1 and Eik is independent (φ = 1or
0, σ = 1 or 0) can be assumed. At this time, the equation (2) representing the disjunctive normal form stochastic logic is equal to the equation for obtaining the true / false value of DNF. That is, the disjunctive standard form of probabilistic logic (extended DNF) is an expression including the DNF of the Boolean function, and the difference between the two is quantitative. Therefore, by selecting the “closest DNF” to a certain extended DNF and predicating this, the extended DN is
The rule represented by F is expressed qualitatively. This is defined as the extended DNF predicate.

If the Euclidean distance is adopted as the distance between the extended DNFs and the distance between the extended DNFs, a certain threshold value (for example, 0.5) is provided and the parameter of the extended DNF is binarized.
NF is obtained. Next, the predication of exception rules will be described. The ALU after learning expresses a rule of P (Eik) → ti (k = 1, 2, ..., K).

From ωbk = -log P (Eik) / | L |, the smaller ωbk, the larger P (Eik).
For the antecedent part of the rule, choose the product of Eik with a small ωbk. This is because ω is binarized by setting an appropriate threshold value, for example, 0.5,
It can be easily done by selecting 0. The consequent part selects -Eo for inhibitory properties and Eo for excitatory properties.

If the same output node has the same γ sign A
When LUs are connected, DNF-type rules can be extracted by treating them as a sum.

However, the rule thus obtained is the extended DN.
Irrespective of F, the product in this case merely indicates the fact that the Eiks happened at the same time, and the sum only indicates that the same kind of ALU was connected. Therefore, in order to distinguish from the rule indicated by the extended DNF, "and *" is used instead of "and" and "or *" is used instead of "or".

When binarizing ωbk, if the inhibitory ALU and the excitatory ALU are very close to each other, different consequent parts may be obtained even if the antecedent part is the same. AL
Basically, all the rules represented by U cannot be ignored, but in this case, the one with a wider firing area, that is, the one with a smaller firing threshold θb is selected, and the ALU with the narrower firing area is discarded.

FIG. 5 is a basic configuration diagram of the fifth invention.

The shaded portion is different from the first, second, third and fourth inventions.

The rule extraction unit 1021 of the fifth invention includes a rule binarization unit 1019 that binarizes the parameters of the parameter unit 1002 of the classification unit 1000 or the exception parameter unit 1010 of the classification device of the second invention. Binarization exception parameter 2
020, and a rule output unit 1020 that outputs the disjunctive standard type classification rule 2021.

First, the binarization threshold 3011 (for example, 0.5) is set in the rule binarization unit 1019, and the rule output unit 10 is set.
In item 20, item names 3012 such as finding items and classification items are set. The rule binarization unit 1019 uses the parameter 2002
By setting the value of (φ, σ) larger than the threshold value 3011 to 1 and setting the smaller value to 0, the binarization parameter 30
10 (φ ′, σ ′) is obtained, and the binarization exception parameter 2020 is obtained from the exception parameter 2005 (ω, γ) in exactly the same manner.
(Ω ′, γ ′) is obtained and sent to the rule output unit 1020. However, in the case of the inhibitory ALU, γ ′ is set to −1 if the absolute value of γ is greater than or equal to the threshold value, and 0 otherwise. Rule output part 1
020 is a binarization parameter 3 corresponding to each classification item
Select the disjunction in which σ'of 010 is 1 and the content is φ'is 1
A general classification rule is predicated by deciding by selecting the finding item of.

For example, regarding the classification item “does not administer insulin”, there are two σs with σ ′ of 1 and one of φ ′ corresponding to one of them is 1 and the finding item is “normal blood glucose level”.
If the other side corresponds to “slightly low body temperature” and “cold outside air”, the predicate is as a rule that “insulin administration is not performed if blood sugar level is slightly normal or body temperature is slightly low and outside air cold”. Be converted.

Further, the rule output unit 1020 determines that γ ′ of the binarization exception parameter 2020 is 0 for each classification item.
Select the one that is not, and make a disjunction, and the content is ω 'where ω'is 0
It is decided by selecting the finding item corresponding to. For example, regarding the classification item “does not administer insulin”, there are two excitatory ALUs, one of which has “high body temperature” as a finding item with ω ′ of 0, and the other has “exercise as a finding item of ω ′ of 0”. If there is one inhibitory ALU, and ω'is 0 as a finding item corresponding to "having a cold", "If you are exercising or * "Do not administer insulin unless you have a high body temperature or a cold".

The rule output unit 1020 uses the parameter 2002
And the rule from the exception parameter 2005 are combined and output as a disjunctive standard type classification rule 2021.
As described above, the function of the rule extraction unit 1021 can be realized.

The threshold 3011 and the item name 3012 are set by the input means 002 by the disjunctive standard type classification rule 2021.
Is output by the output means 003.
9. The rule output unit 1020 can be realized by the computer 001. Next, the means for realizing the classifying device according to the sixth aspect of the present invention will be described, but before that, the method for recalling input will be described.

The recall process outputs the output squared error under the constraint of 0 ≦ Pik ≦ 1.

[0153]

[Formula 19]

Is to obtain Pik (k = 1, ..., K) that minimizes.

This processing can be realized by the non-linear optimization method described in the rule learning section 1012. An example of the correction amount calculation of the input probability 2001 when this minimization is performed by the descent method is shown below.

[0156]

[Equation 20]

FIG. 6 is a basic configuration diagram of the sixth invention.

The shaded portion is different from the first, second, third, fourth and fifth inventions.

First, the target output 2022 is input to the target output setting unit 1022, and the allowable recall error 3 is input to the recall convergence determination unit 1026.
013 is input to the input holding unit 1023 and the input probability initial value 301
Set 4. This processing can be realized by the input means 002.

First, the input holding unit 1023 sets the input probability initial value 3014 to be held as the input probability 2001 to the first value.
Sent to the classifying device of the invention.

The recall error calculation unit 1024 calculates a recall error 2023 (corresponding to − (ti-si)) from the output probability 2003 of the first classifier and the target output 2022, and the input correction unit 1025 and the recall convergence determination. Send to section 1026.

The input correction unit 1025 obtains the input correction amount 2024 according to the equation (13) and sends it to the input holding unit 1023.

The input holding unit 1023 stores the input correction amount 202
4, the input probability is corrected by adding it to the retained input probability 2025 held.

The recall convergence determination unit 1026 determines the recall error 20.
If 23 is larger than the allowable recall error 3013, the recall continuation signal 2027 is sent to the input holding unit 1023, and the input probability held by the input holding unit 1023 is sent again to the first classification device. When the recall error 2023 is smaller than the allowable recall error 3013, the recall continuation signal 2027 is not sent but the output request signal 2028 is sent to the recall probability output unit 1027.

The recall probability output unit 1027 outputs the output request signal 2
Upon receiving 028, the holding input probability 2025 is subjected to processing such as graphing and output as a recall result 2026.

With the above, recall of the input value can be realized.

Objective output setting unit 1022, recall error calculation unit 1024, input correction unit 1025, recall convergence determination unit 102
6 is a computer 001 that causes the recall probability output unit 102
7 can be realized by the output means 003.

[0168]

The classifying device using the disjunctive normal form stochastic logic of the present invention has a classification rule (or a diagnosis rule, a judgment rule).
Can be retained as a rule expression that guarantees quantitativeness and is easy for humans to understand, and the classification process can be performed semi-automatically based on this, and existing classification rules can be converted to the rule expression and used. It holds exceptional classification rules such as individual differences and can classify using them, and it can learn the rule expressions semi-automatically using classification cases, and it can semi-automatically learn exceptional classification rules from exceptional classification cases. Since the classification rule held by the device can be extracted as a rule that can be understood by humans and the target value of the input can be reversed from the target output, diagnosis in the medical and medical engineering fields and treatment / control based on it can be performed. I can help.

[Brief description of drawings]

FIG. 1 is a basic configuration diagram of a first invention.

FIG. 2 is a basic configuration diagram of a second invention.

FIG. 3 is a basic configuration diagram of a third invention.

FIG. 4 is a basic configuration diagram of a fourth invention.

FIG. 5 is a basic configuration diagram of a fifth invention.

FIG. 6 is a basic configuration diagram of a sixth invention.

[Explanation of symbols]

 001 computer 002 input means 003 output means 1000 classification unit 1001 input unit 1002 parameter unit 1003 output probability calculation unit 1004 output unit 1005 probability conversion unit 1006 knowledge conversion unit 1007 rule analysis unit 1008 rule setting unit 1009 exception correction unit 1010 exception parameter unit 1011 Exception determination unit 1012 Rule learning unit 1013 Error calculation unit 1014 Parameter correction unit 1015 Convergence determination unit 1016 Exception error calculation unit 1017 Exception parameter correction unit 1018 Exception convergence determination unit 1019 Rule binarization unit 1020 Rule output unit 1021 Rule extraction unit 1022 Purpose Output setting unit 1023 Input holding unit 1024 Recollection error calculation unit 1025 Input correction unit 1026 Recollection convergence determination unit 1027 Recollection probability output unit 2000 Measured value 2001 Input probability 2002 Parameter 2003 Output probability 2004 Classification result 2005 Exception parameter 2007 Exception signal 2008 Error 2009 New parameter 2010 Classification case 2011 Correction stop instruction 2012 Classification result probability 2014 Exception classification case 2015 Exception input probability 2017 New exception parameter 2018 Exception classification result probability 2019 Exception error 2020 Binarization exception parameter 2021 Disjunction standard type classification rule 2022 Objective output 2023 Recall error 2024 Input correction amount 2025 Retention input probability 2026 Recall result 2027 Recall continuation signal 2028 Output request signal 3000 Existing classification rule 3002 Modified output probability 3003 Allowable Error 3004 Error initial value 3005 Case presentation end signal 3006 Exceptional allowable error 3007 Exceptional error initial value 3008 Calculation in-progress result 3009 Correction end Signal 3010 Binarization Parameter 3011 Threshold 3012 Item Name 3013 Allowable Recall Error 3014 Input Probability Initial Value

Claims (6)

[Claims] 1. An input unit for inputting an input probability, a parameter unit for holding a parameter of a rule expressed by a disjunctive normal form type probability logic, and an output probability for calculating an output probability from the parameter and the input probability. It has a classification unit consisting of a calculation unit and an output unit that outputs the output probability as it is or after processing it, and a rule setting unit that sets existing classification rules and a disjunctive standard form that analyzes the set rules. Classification having a knowledge conversion unit including a rule analysis unit that obtains a parameter of a stochastic logic of a type and sets it in the parameter unit, converts a measurement value into an input probability, and inputs the probability to the classification unit. apparatus. 2. An exception parameter part that holds a parameter of an exception rule that takes precedence over a given classification rule, and an exception judgment that, when an input is obtained, determines whether or not it satisfies the antecedent part of the exception rule. Section, and if it satisfies, the output probability is corrected from the input probability and the parameter of the exception rule and passed to the output section, and if not satisfied, the output probability is not corrected and passed to the output section The classification device using the probabilistic logic of the disjunctive normal form according to claim 1, further comprising a section. 3. An error calculation unit that obtains an error between the output probability output by the classification unit and the classification result probability in the classification data by using the input probability in the classification case composed of the input probability and the classification result probability, and the error. Based on the error calculated by the calculation unit, a parameter correction unit that corrects the parameter of the parameter unit in the direction of decreasing the error, and a convergence determination that repeats the parameter correction until the error becomes equal to or less than a preset allowable error. 3. A classification device using the disjunctive normal form stochastic logic according to claim 1 or 2, further comprising a rule learning unit including a section. 4. An exception error calculation for calculating an error between the output probability from the exception correction unit and the exception classification result probability by using the exception input probability in the exception classification case consisting of the exception input probability and the exception classification result probability. Section, an exception parameter correction section that corrects the parameters of the exception parameter section based on the exception error calculated by the exception error calculation section so that the exception error decreases, and the exception error becomes less than or equal to a preset allowable error. The classification device using the disjunctive standard logic according to claim 2, further comprising: an exception rule learning unit including an exception convergence determination unit that repeatedly corrects the exception parameter. 5. A rule binarization unit that binarizes the parameter unit of the classification unit or the parameter of the exception parameter unit, and outputs a disjunctive normal form type classification rule from the binarized parameter. The classification device using the probabilistic logic of the disjunctive normal form according to claim 1 or 2, further comprising a rule extraction unit including a rule output unit. 6. A target output setting unit that sets a target output probability in order to inversely estimate an input probability necessary to obtain a target output probability using the classification unit, and holds the input probability. An input holding unit, a recall error calculation unit that obtains an error (recall error) between the output probability output by the classifier and the target output probability using the input probability of the input holding unit, and a direction to reduce the recall error. An input correction unit that corrects the input probability held by the input holding unit, a recall convergence determination unit that repeatedly corrects the input until the recall error is equal to or less than a preset allowable recall error, and the recall error is less than the allowable recall error. 2. The classification device using the disjunctive standard form stochastic logic according to claim 1, further comprising a recall probability output unit that outputs the input probability held by the input holding unit.