JP2632117B2 - Case-based reasoning device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a case in which past empirical cases are stored in a case database, and existing cases similar to the conditions of the new case are selected and corrected to infer the conclusion of the new case. It relates to a base inference device.

[0002]

2. Description of the Related Art Case-based reasoning technology is based on the
This is a technique for generating an inference model based on the hypothesis that "it is performed based on past experience." Basically, the following inference is performed: First, the cases that have been experienced in the past are accumulated and the case base is created. When the conditions of the new case are input, the existing case most similar to the condition is searched from the case base, and the searched existing case is appropriately corrected to infer the conclusion of the new case. In addition, learning of new cases updates the case base.Inference technology such as this is applicable to the field of so-called natural language processing, which handles individual data such as characters and words as examples. Various methods have been proposed for realizing an inference apparatus in this field.

[0003]

However, such a case-based reasoning technique is used for predicting the load of a district heating and cooling system.
When trying to apply to technologies such as control of the valve operation amount of a plant and identification of equipment characteristics, in the above systems and equipment, cases are not individual data such as characters and words, but are observed in a time series and continuously. Since the data is continuous data such as a changing temperature, the following problem occurs.
That is, the first problem is how to represent the case of continuous data to create a case base,
The problem is how to determine the similarity between continuous data cases and search for similar cases for new cases. The third point is how to correct similar cases in continuous data to conclude new cases. The fourth problem is how to learn a new case of continuous data and update the case base. In addition, the conventional case base is a technology having a huge storage capacity and a large amount of calculation, such as having 32k processors. There is also a problem that it is difficult to realize “execution of processing”. Therefore, an object of the present invention is to enable a case of continuous data such as temperature to be processed by a small-scale controller or the like.

[0004]

SUMMARY OF THE INVENTION In order to solve such a problem, the present invention is an inference apparatus for inferring a case in which continuity is established between each case. Discretize the value of each input variable, symbolize the condition part with these discrete values, integrate the conclusion part of the existing case for the same symbol, integrate this conclusion part with the condition part, and change the conclusion part with respect to the change of the condition part A case base generation unit that generates information along with the information; and a means for determining a threshold used to determine a similarity of a condition part in consideration of inference accuracy from existing cases as preprocessing for inference; A parameter determination unit comprising means for determining a restriction on the number of similar cases used for inference, a unit for determining the similarity of an existing case to a condition part of a new case based on the continuity of the phase of the condition part, and a condition part Based on the similarity of A means for searching for similar cases to the new case from the case base in consideration of the degree of influence of the conditional part on the conclusion part, a means for determining the importance of similar cases for the new case from the distance of the conditional part based on the phase, Means for correcting and integrating multiple similar cases based on the change in the conclusion part with respect to the change in the condition part to infer the conclusion part of the new case, and the condition part of the similar case based on the continuity of the condition part and the conclusion part If the phase of the conclusion is close, the phase of the conclusion part is also close, and the case-based reasoning section consisting of means for determining the credibility of the inference result and means for outputting information about the inference result and its credibility, and the new case as it is And a case base learning unit for learning in real time and updating the case base.

[0005]

[Operation] When a new case is input, a case to be used for inference of a conclusion part of the new case is selected from existing cases, and a similarity is calculated, and a distance between the similar case and the new case based on the similarity is calculated. At the same time, the importance based on the distance is calculated, and the new case is inferred in consideration of the importance. When a similar case used for the inference is searched, the credibility is calculated based on the similarity.

[0006]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing one embodiment of a system to which a case-based inference device according to the present invention is applied. This system is an air conditioning load prediction system for district cooling and heating.
In the figure, reference numeral 1 denotes a storage device for storing case data such as an outside air temperature observed at certain time intervals in the past and a discomfort index at this time, and 2 denotes a data classification unit for classifying the case data for each month. Numeral 10 creates a recognition model from the case data for each month, and when new data such as outside temperature is actually input, recognizes the attribution to each month based on this recognition model and performs load prediction. The state recognition automatic generation device 20 creates a causal relationship model for each month from the case data for each month, and when new data is input, makes inferences based on the causal relationship model and calculates its credibility. An inference apparatus 30 is a load predicted value display unit, 40 is a credibility integration unit, and 50 is a credibility display unit.

Here, the state recognition automatic generation device 10 includes a recognition model generation unit 11, a recognition model learning unit 12, a recognition model storage unit 13, a recognition model situation input unit 14, an attribution recognition unit 15, and a final predicted value determination unit. 16. Then, the recognition model generating means 11 extracts factors for recognition from past case data to generate a recognition model based on each factor, and the recognition model learning means 12 generates a recognition model when inputting new case data such as outside air temperature. And updates the above recognition model, and the like, and these recognition models are stored in the recognition model storage unit 13. On the other hand, when the new case data is input from the recognition model situation input unit 14, the attribution recognizing unit 15 recognizes the attribution to each factor based on the recognition model in the storage unit 13, and finalizes the recognition result. In addition to sending the result to the value determining means 16, the final predicted value determining means 16 integrates the recognition result and the inference result of the inference apparatus 20 and outputs the result as a final load predicted value.

The inference apparatus 20 includes a causality model generation means 21, a causality model learning means 22, a causality model storage unit 23, a causality model situation input unit 24, an inference means 25, and a credibility calculation means 26. Have been. Then, the causal relation model generation means 21 generates a model of a causal relation between the input factor and the result several hours after the past case data, for example, room temperature, and the causal relation model learning means outputs the new case data. When this is input, it learns this and updates the above causal relationship model, etc.,
These causal relationship models are stored in the causal relationship model storage unit 23. On the other hand, when the new case data is input from the causal relationship model situation input unit 24, the inference means 25 makes an inference based on the causal relationship model in the storage unit 13 for this new case data, and outputs the inference result to the final prediction value determination. The load is sent to the means 16 and integrated with the recognition result by the automatic state recognition generation device 10 to display the predicted value on the load predicted value display section 30. Further, the inference means 25 sends the inference result to the credibility calculation means 26 to cause the credibility calculation unit 26 to perform credibility calculation and display the result on the credibility display unit 50. When inferring new case data, the inference means 25 makes an inference based on the topology. The phase refers to a structure given to a set so that the concept of continuity can be defined, and indicates, for example, a distance between similar new case data and past case data, similarity between similar cases, and the like.

Here, the inference apparatus 20 is, as shown in FIG. 3, hardware-based, in which existing case data (past case data) is used.
, An online input device 62 for inputting new case data, an algorithm storage unit 63,
It comprises a CPU 64 for performing an inference operation in accordance with a predetermined algorithm stored in an algorithm storage unit 63, a data storage unit 65 for storing various data, and an output device 66, and is functionally shown in FIG. It consists of each means.

That is, the above-described causal relation model generating means 2
1, a memory base creating unit 21A, a parameter determining unit 21B, the causal relationship model learning unit 22,
The inference unit 25 includes a similarity determination unit 25A, a similarity case search unit 25B, an importance determination unit 25C, a case integration unit 25D, the credibility calculation unit 26, and an output unit 27.

As described above, the inference apparatus 20 deals with the causal relationship between the input factor of the past case data and the result several hours later, for example, the room temperature, and X = <x1, x2, ..., xn>
And Y = <y> as the output space,
Events X1 (t), X2 (t),.
.., Xn (t) produce output Y (t + α) after α hours, ie, {X1 (t), X
2 (t),..., Xn (t), Y (t + α)｝ (t =
1,..., N) and infers data in which each input / output variable changes continuously.

In such an inference apparatus, a case base is first created for past case data by the causal relation model means 21 before inference is performed. That is, an input space is first discretized and divided into a finite number of input events, and one case is generated by integrating input / output data belonging to the same input phenomenon. At this time, the condition part of the case, that is, the data such as the outside air temperature and the discomfort index is obtained by discretized input data ｛X
1, X2,..., Xn}, and the conclusion of the case,
That is, the room temperature after α hours is the barycentric value Y of the output data, the number n of times the same input event has occurred, and the barycentric value Δ
.., ΔY / ΔXn, that is, ｛Y, n, Δ
Y / ΔX1,..., ΔY / ΔXn}.

For example, as input data, the outside air temperature X1
(° C.) and the discomfort index X2 (%) are observed, and the maximum value (max) and the minimum value (min) of these data during each time t (t = 1,..., N) in a certain period are obtained. each, X1 (max) = 30.0, X1 (min) = 20.0 X2 (max) = 80.0, X2 (min) = 70.0 , and the and the data at time t = t ₁ X1 (t ₁ ) = 25.6, X2 (t ₁ ) = 78.7, and the discretization number for discretizing this input space is “10” (division between the maximum value and the minimum value is divided into 10). Then, the discretized data is, for example, X1 = 6, X2 =
9 and is symbolized by one event {6,9}. Here, the room temperature Y (t at time t = t ₁ + α)
_{If (1} + α) is 25.0, a causal relationship of {6, 9} → 25.0 is obtained.

If the event at time t = t ₁ +1 is X1 (t ₁ ) = 25.8, X2 (t ₁ ) = 78.79 and room temperature Y (t ₁ + α + 1) = 25.5, The input events belong to the same {6, 9}, and {6, 9} → 25.25 [= {25.0 + 25.5} /
2, where n = 2] and the case data is compressed. As a result, the capacity of the memory required for the case base can be significantly reduced as compared with the conventional example. Further, each partial differential value described above is
This is the amount of change in the output with respect to the amount of change in each input variable. In this case, since each input / output variable is continuous data, the partial differential value ΔY / ΔXi (t) can be calculated by equation (1).

[0015]

(Equation 1)

As described above, after a case base is created for a past case (existing case), inference about a new case is performed by the inference means 25. First, the condition part of the new case is {Xi ^* } (i = 1, 2,..., N), and the existing case is {Xi, Y, n, ΔY / ΔXi} (i = 1,
2,..., N). Here, the condition part of the new case is
At the same time as the input, it is discretized and symbolized as described above.

Here, when inferring a new case, first, the similarity of the existing case to the new case is determined by the similarity determining means 25A.
(This similarity corresponds to the concept of a neighborhood system in phase). The similarity of the existing case to the new case is determined by the following definition. That is, the similarity 0 is | Xi ^* −Xi | = 0 (i = 1,
2,..., N) The similarity 1 is | Xi ^* −Xi | ≦ q _Xi (i = 1,
2,..., N) Similarity 2 is | Xi ^* −Xi | ≦ q _Xi +1 (i = 1,
2,..., N) The similarity 3 is | Xi ^* −Xi | ≦ q _Xi +2 (i = 1,
2,..., N). Here, q _Xi is called a threshold,
It is a digit value determined from variance of Xi (condition part of existing case) with respect to allowable accuracy of Y (conclusion part of existing case) from existing case data.

As an example, when X1 is the outside air temperature, X2 is the discomfort index, and the room temperature after Y = α hours is considered, when an error between the room temperature Y and the predicted value is attempted within 2 degrees, an input is made. Discretization is performed so that one digit becomes twice as in the case of space. If the discretization number is, for example, "20", Y ^* (conclusion part of a new case) belonging to the same digit for each of digits 1 to 20 is collected from existing case data, clustered, and belong to the same class. The variance of the conditional parts X1 and X2 of the new case is obtained. The variance of each cluster of the condition parts X1 and X2 of the new case belonging to the digit i of Y (conclusion part of the existing case) is calculated as a discrete value, and averaged by the number of clusters to obtain digit values qi _X1 and qi _X2 . . Digit values qi _X1 and qi _X2 are averaged according to the following equation. That is, q _X1 = Σqi _X1 / 20, q _X2 = Σqi _X2 / 20

Here, the closest digit value is q _X1
= 2, q _X2 = 3 (actually, q _X1 = 2.123...,
q _X2 = 3.456...). However, if Y needs to be within 1 degree, q _X1 = 1 and q _X2 = 2, and the digit value q _Xi differs depending on the required precision. That is, the distance between the existing case and new case for each variable Xi is the phase of the condition part is smaller than q _Xi is closer, the conclusion part of the new case at that time is required for the conclusion of the existing case Considered to be within precision.

Next, a similar case to the new case is searched by the similar case search means 25B. Existing cases of the optimal case are extracted as similar cases in the order of similarity to the new case. As the optimum number of cases, for example, the number of cases that makes the best inference from a simulation based on existing cases is selected.
If there are more existing cases with the same similarity than the number of optimal cases, the degree of influence of each variable Xi on Y, that is, the correlation coefficient R
_The priority is set for each variable according to the magnitude of _Xi and extracted.

Next, the importance determining means 25C determines the importance of a similar case with respect to the new case. A distance is defined in the input space and the phase between cases is considered. Here, as an example, a distance L as shown in Expression (2) is introduced.

[0022]

(Equation 2)

Here, Φi is the weight of the distance in the variable Xi. Then, the importance Wj at the time of inference of the extracted m similar cases is defined using Expression (3). That is,

[0024]

(Equation 3)

FIG. 5 is an explanatory diagram for explaining the importance of such similar cases, that is, the situation before and after weighting of similar cases. A new case Xi ^* (i = 1, 2,..., N) is obtained using m similar cases extracted in this way, the number of which is m.
Calculated integrating using an inference value Y ^* to (4) for (case integration means 25D).

[0026]

(Equation 4)

Where Lij is the distance of the i-th case from the input data on the j-input variable axis, yi is the conclusion value of the i-th similar existing case, and ΔY / ΔXj is j of the i-th similar existing case. The percentage of the change that the second change gives to the conclusion value is shown.

Next, the credibility of the inference result is determined based on the calculation result of the credibility calculating means 26. That is, the credibility of the inference result is determined using the similarity of the similar case used in the inference to the new case. For example, if the similarity used in the inference is similar to the new case, the highest similarity is defined as the credibility of the inference result.
The inference result with the highest similarity “1” is determined to have credibility “1”. In this case, the credibility “0” has the highest credibility, and the credibility decreases as the number increases. The output means 27 can output the obtained inference result and its credibility, and can also output the similar case used for the inference.

Next, when performing case-based learning, a new case is learned by the causal relationship model learning means 22 to update the case base. Such case-based updating is performed by the following procedure. However, those marked with * indicate new cases. That is, assuming that the number of events of the same condition part up to the previous time is n, the number of events is increased by one and n + 1
And the output value Y is (Y × n + Y ^* ) / (n +
1), and the partial differential value ΔY / ΔX1 is set to (ΔY / ΔX
1 × n + ΔY / ΔX1 ^* ) / (n + 1).

FIG. 4 is a flowchart summarizing the operation of the inference means 25 for performing the above inference. That is, when new case data X such as outside air temperature is input as a condition part of a new case, a case to be used for inference of room temperature after a predetermined time, which is a conclusion part of the new case, is selected from existing cases. The similarity between the case and the existing case is calculated, and a similar case is searched (step ST1). Then, the distance between the searched similar case used for inference and the new case is calculated, and the importance based on the distance is calculated (step ST).
2), the conclusion of the new case is inferred in consideration of the importance (step ST3), and the inference result y is obtained (step ST4). On the other hand, if a similar case used for inference is found, its credibility is calculated based on the similarity (step ST
5) Obtain the result (step ST6). Then, the credibility result and the inference result y are output (step S
T8).

When the fact Y (the conclusion of the new case) is obtained, the fact Y is compared with the inference result y (step ST9). Then, the fact Y differs from the inference result y,
If the new case X does not exist in the past case, X, which is the condition part of the new case, and Y, which is the conclusion part thereof, are registered as new cases (step ST10).

[0032]

As described above, according to the present invention,
When a new case is input, a case to be used for inference of the conclusion of the new case is selected from the existing cases, the similarity is calculated, the distance between the similar case based on the similarity and the new case is calculated, and the distance is calculated. Is calculated, and the new case is inferred while considering this importance.When similar cases used for the inference are searched, the credibility is calculated based on the similarity. Indispensable for predicting, controlling, and identifying complex systems that change continuously, high-precision inference, credibility evaluation of inference results, and real-time adaptive learning are possible, and case data is symbolized. As a result, the memory capacity and calculation amount are reduced, and as a result, real-time processing can be executed by a small-scale controller.

[Brief description of the drawings]

FIG. 1 is a functional block diagram showing an embodiment of a system to which a case-based inference device according to the present invention is applied.

FIG. 2 is a functional block diagram of the inference apparatus.

FIG. 3 is a block diagram of the inference apparatus.

FIG. 4 is a flowchart showing an operation of the inference apparatus.

FIG. 5 is a diagram showing a weighting situation by calculating importance of similar cases in the inference apparatus.

[Explanation of symbols]

 21 Causal relation model generating means 21A Memory base generating means 21B Parameter determining means 22 Causal relation model learning means 25 Inference means 25A Similarity determining means 25B Similar case searching means 25C Importance determining means 25D Case integrating means 26 Credibility calculating means 27 Output means

Claims (1)

(57) [Claims] An inference apparatus for inferring a case where continuity is established between each case, discretizing the value of each input variable of a condition part of an existing case and symbolizing the condition part by the discrete value And a case base generation unit that integrates the conclusion part of the existing case for the same symbol, generates the relation between the conclusion part and the condition part, and generates the change information of the conclusion part with respect to the change of the condition part. Means for determining a threshold used to determine the similarity of a condition part in consideration of inference accuracy from existing cases as preprocessing for
A parameter determination unit comprising means for determining a restriction on the number of similar cases used for inference, a unit for determining the similarity of an existing case to a condition part of a new case based on the continuity of the phase of the condition part, and a condition part Means to search for similar cases to a new case from the case base based on the similarity of a condition, taking into account the degree of influence of the conditional part on the conclusion part, and determine the importance of the similar case to the new case from the distance of the conditional part by phase Means,
Based on the importance, multiple similar cases are modified in consideration of the change of the conclusion part with respect to the change of the condition part, integrated, and the means to infer the conclusion part of the new case and the continuity of the condition part and the conclusion part A case-based reasoning unit comprising means for determining the credibility of the inference result by using the fact that the phase of the condition part is close to the phase of the conclusion part, and a means for outputting the inference result and information on its credibility; And a case base learning unit for updating the case base in real time as it is.