JPH09297685A - Action deciding method for inference device, and device for executing the method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a decision result equivalent to that obtained when a human being describes the loss value by securing the correspondence between the proposition of an action pair and the small real number value when the similarity is high to the optimum action pair, leading the expected loss value of an action based on the loss value and the posterior probability value, and selecting an action of the least expected loss value. SOLUTION: A loss data processing part 2 sends a pair of propositions to a similarity calculation part 3 to secure the correspondence between the propositions and the similarity of >=0.0 and <=1.0 real number value. A loss value calculation part 4 secures the correspondence between an optional action pair and the loss value of >=0.0 and <=1.0 real number value. At the same time, the correspondence is secured between the optimum action pair and 0.0. Then the correspondence is secured between the loss value and the small value when the similarity is high to the proposition of the optimum action pair. A probability calculation part 6 leads the posterior probability value of each proposition from a fact and the probability data to output it to an expected loss value calculation part 7, and the expected loss value of each action is led from the posterior probability value and the loss data given from the part 2. Then an action decision part 8 extracts an action of the least expected loss value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an action determining method in an inference apparatus and an apparatus for implementing this method,
In particular, in an inference device of an automatic device having a knowledge processing unit such as an expert system or an intelligent robot, an appropriate solution for a user is selected from a plurality of solution candidates, and the purpose of the automatic device is selected from a plurality of action patterns. The present invention relates to a method for determining an action in an inference device that selects an action suitable for, and a device for performing this method.

[0002]

2. Description of the Related Art A conventional example will be described with reference to FIGS. An inference device that gives advice to a human action needs a function of selecting a proper action for a human from a plurality of actions. As a method to realize such a function, "action to reduce the degree of loss human suffers as much as possible"
Is adopted as an appropriate action. Consider a case where a human being asks the reasoning device about “clothes or belongings when going out”. The weather gauge has {high}, {medium}, and {low} scales, the weather has {fine}, {rain}, and {snow}, and the actions are {hands}, {umbrella}, and {umbrella}. And boots}
Suppose that there is each. The questioner is a fact {weather gauge:
When High} is input, the inference apparatus first obtains the posterior probability value of each weather corresponding to the scale of the weather meter from the probability data P shown in FIG. In FIG. 3, P (clear | high) = 0.
7 indicates that when the scale of the weather gauge indicates {high}, the probability of clear weather is 0.7. In this example, the posterior probability values of each weather {clear}, {rain}, and {snow} can be acquired as 0.7, 0.2, and 0.1 based on FIG. Furthermore, the loss value of each action is acquired from the loss data L shown in FIG. In FIG. 4, L (rain, empty-handed) = 1.0
Is a value that evaluates the degree of loss when it goes out by hand and it rains. In this case, the maximum loss value is 1.0 and the minimum value is 0.0, and the degree of loss is expressed as a real value between 0.0 and 1.0. Therefore, a loss value of 1.0 means that the loss is very large. Then, the inference apparatus calculates the expected loss value which is the expected value of the loss value for each action from the posterior probability value P and the loss value L thus acquired. For example, the expected loss value when taking an action {umbrella} with a weather meter {high} is L (fine, umbrella) x P (fine | high) + L (rain, umbrella) x P
(Rain | High) + L (Snow, Umbrella) x P (Snow | High) = 0.5 x
It is calculated as 0.7 + 0.0 × 0.2 × 2 + 0.3 × 0.1 = 0.38. This expected loss value is a value obtained by calculating the loss value predicted for each action when only the scale of the weather gauge is known. Similarly, the expected loss values of each action {empty hand} and {umbrella and boots} when the weather meter is {high} are calculated to be 0.3 and 0.46, respectively. The action {empty} that minimizes the expected loss value calculated in this way to 0.3 means that it is the action that minimizes the predicted loss value. You can Therefore, the inference apparatus selects an appropriate action {blank} and presents the action with the smallest expected loss value.

In order to realize a system for responding to various worries or hesitations held by humans based on such a method, it is necessary to preliminarily incorporate loss data for many propositions and actions into an inference device. Since the number of loss values required for inference is “the number of actions × the number of propositions” as shown in FIG. 4, if the number of events and actions handled by the inference apparatus is large, the number of loss values required will be these. As a result, the number of loss values increases dramatically, and as a result, it becomes necessary to incorporate a huge number of loss values.

In order to obtain the loss data, a method in which a person directly describes a numerical value has been adopted. For example, the loss value when the behavior is {hands up} and the weather is {rain} is considered to be large because the clothes are soaked, and the loss value is described as a large value 1.0. On the other hand, the loss value when the behavior is {umbrella} and the weather is {clear} is less than 1.0 because it is not so good but it is better than soaking wet, so it is described as 0.5, for example. However, such a loss data acquisition method based on "quantification work that humans are not good at" is a difficult method to obtain a large amount of loss data that can respond to various worries or hesitations of humans. . Such difficulty limits the problem that the reasoning apparatus can handle to a specific narrow area, and causes an undesirable restriction.

[0005]

As described above, human beings subjectively describe all loss values,
It is difficult to construct an inference device that can handle large-scale problem areas. The present invention is a behavior determination in an inference device that can obtain a determination result equivalent to that when a human describes a loss value by automatically generating a loss value, even if the human does not describe all loss values. A method and an apparatus for carrying out the method are provided.

[0006]

[Means for Solving the Problems] Regarding a plurality of propositions whose values are probability values, a certain fact, probability data that derives the posterior probability value of the proposition based on the facts, and an action set that combines one action with the proposition In the action determination method in the inference apparatus that selects one action using the optimal action set in the above, the proposition of any two action sets is associated with a similarity degree consisting of a real value of 0 or more and 1 or less, 0 or more 1 for any one action group
When associating with a loss value consisting of the following real values, the optimal action set is associated with 0, and the action set of the optimal action set and the proposition whose similarity is s are small values when s is large. Inference that associates with a large value when s is a small value, derives the expected loss value of each action based on the loss value and the posterior probability value, and selects one action with the smallest expected loss value. The behavior decision method in the device was constructed.

In the behavior decision method in the inference apparatus, the behavior decision method in the inference apparatus is the point of view similarity, which is the similarity between two propositions when the behavior is viewed from the viewpoint. The loss data processing unit 2 is provided with the loss data processing unit 2 to which the optimum action set (proposition W, action A) and the action set (proposition X, action A) are input.
The similarity calculator 3 that calculates the similarity between the propositional set X and the propositional set (proposition W, propositional X) of the action set having the action A given from A and returns the calculation result to the loss data processing unit 2 is provided. , A loss calculation unit 4 for calculating the loss value of the action set based on the similarity of the propositional set passed from the loss data processing unit 2 and returning the calculation result to the loss data processing unit 2 The expected loss based on the posterior probability value and the loss data, which is a set of the loss values of the action set returned to the loss data processing unit 2, is provided with the probability calculator 6 that calculates the posterior probability value of the proposition based on the data. The action determination device in the inference device is configured to include the expected loss calculation unit 7 that calculates data, and the action determination unit 8 that extracts the minimum expected loss value from the input expected loss data.

Further, in the action determining device in the above inference device, the inference device in which the similarity in the similarity calculating unit 3 is converted into the viewpoint similarity which is the similarity between two propositions when the action is viewed from the viewpoint. The behavior decision device in.

[0009]

BEST MODE FOR CARRYING OUT THE INVENTION Here, terms frequently used in the present invention are collectively defined. Action set: One proposition and one action set Loss value of action set: Loss value when the action of the action set is taken on the truth that is the probability value 1 of the proposition of the action set Loss data: All actions A set of loss values for a set Propositional set: a set of two propositions Similarity of a propositional set: similarity of two propositions of a propositional set This invention is based on the degree of similarity between two propositions, that is, the degree of similarity. Even if a loss value is automatically calculated, a behavior determination method and device that obtain a determination result equivalent to that when a human directly describes the loss value is realized.

Now, consider three loss values of an action group (fine, empty-handed), an action group (rain, empty-handed), and an action group (cloudy, empty-handed). First, when the weather, which is a proposition, is fine,
No umbrellas or boots are needed, so there is no loss when going out empty-handed, and the loss value of the action group (fine, empty-handed) is set to 0.0. On the other hand, the loss value of the action group (rain, empty-handed) is large because it gets wet when it goes out empty-handed. In this way, the loss value of the action group (rain, empty-handedness) changes greatly compared to the loss value of the action group (fine, empty-handedness), depending on whether "fine" and "rain" drain water or not. This is due to the fact that the weather is essentially different. That is, the weather {rain} and the weather {clear} are completely different, and when the action {hand-held} is taken, the weather is {clear}.
Depending on whether it is rain or not, the loss value that human suffers will change greatly. Thus, it can be seen that the loss value changes significantly when the weather is not similar.

Next, if it becomes cloudy when you go out with your hands empty, there is a loss that you will be anxious because it is likely to rain, but since it does not get wet, the loss value of the action group (cloudy, empty-handed) is very small. become. In this way action groups (cloudy,
The reason why the loss value of (hand-held) does not change much compared to the loss value of the action group (fine, hand-held) is that "fine" and "cloudy" are essentially similar in terms of whether or not to drain water. Because it is the weather. Thus, it can be seen that when the weather is similar, the loss values are also similar and do not change much.

As described above, the present invention employs the phenomenon that "the loss value changes little if the propositions are similar", and a determining method and apparatus for selecting an appropriate action based on the loss value calculated. To realize. The reasoning device has an action set (proposition X, action A) for an optimum action set (proposition W, action A).
Is calculated as "a small value if the degree of similarity of the set of proposition X and proposition W is large" and "a large value if the degree of similarity is small". Loss data is obtained by applying this calculation method to all optimal behavior sets. The present invention utilizes the loss data thus calculated. When the questioner gives a set of facts, the inference device generates posterior probability data of each proposition based on the probability data. Then, expected loss data is generated from the posterior probability data and the loss data generated as described above. Select one of the smallest expected loss values.

To explain the present invention in more detail, the similarity between the proposition W and the proposition X is calculated for the optimum action set (proposition W, action A) and the action set (proposition X, action A). If the degree of similarity is large, it means that the proposition W and the proposition X are similar to each other, so that the loss value of the action set is not so different from the loss value of the optimum action set. For example, considering the optimum action group (fine, empty-handed) and the action group (cloudy, empty-handed), “sunny” and “cloudy” are similar weathers in that moisture does not fall, and therefore the similarity is set to a large value. It is calculated. Therefore, the optimal action group (fine, empty-handed) and the action group (cloudy,
The loss value of the empty hand will not change much. That is,
It can be seen that the loss value of the optimum action group (fine, empty-handed) is small, so that the loss value of the action group (cloudy, empty-handed) is also small. As described above, the loss value of the action set (proposition X, action A) is calculated as a small value.

On the contrary, if the degree of similarity is small, it means that the proposition W and the proposition X are not similar. Therefore, the loss value of the action set (proposition X, action A) is the optimal action set (proposition W, action). Compared with the loss value of A), it will change greatly. For example, when considering the optimum action group (fine, empty-handed) and the action group (rain, empty-handed), “sunny” is a different weather in that it does not get water as compared with “rain”, and therefore the similarity between the two is similar. Is calculated as a small value. Therefore, the loss values of the optimum action group (fine and empty-handed) and the optimum action group (rain and empty-handed) greatly change. Here, it can be seen that the loss value of the optimum action group (fine, empty-handed) is small, and the loss value of the action group (cloudy, empty-handed) is large. As described above, the loss value of the action set (proposition X, action A) is calculated as a large value.

By calculating the loss value in this way, the present invention can approximately obtain the loss value based on human subjectivity. The expected loss value calculated from the loss data thus obtained approximately gives the expected value of the loss value that the questioner will suffer. Therefore, by presenting the action that minimizes the expected loss value, it is possible to achieve the object of the present invention to select and present one solution that is appropriate for the questioner.

[0016]

An embodiment of the present invention will be described with reference to FIG. First, a set of optimum action sets and a set of action sets are input to the loss data processing unit 2 to execute a process of generating loss data. The loss data processing unit 2 generates a loss value for all action sets including the action A based on a certain optimum action set (proposition W, action A). More specifically, a set of proposition X and proposition W is passed to the similarity calculation unit 3 to calculate the similarity, and the loss data processing unit 2 receives the similarity of the proposition set. Further, the degree of similarity is passed to the loss calculation section 4 to calculate the loss value, and the loss data processing section 2 receives the loss value. The loss value obtained as described above is set as the loss value of the action group (proposition X, action A). When the loss value calculation is completed for all the action groups including the action A, the loss data processing unit 2 executes the same calculation for other actions other than the action A. When the loss values have been obtained for all the action groups, the calculation processing operation of the loss data processing unit 2 ends. The loss data can be obtained as described above. here,
When the probability calculator 6 is given a fact, that is, the state of the weather meter, a process of deriving the posterior probability value of each proposition from this fact and the probability data is executed. The posterior probability value of each proposition derived in the probability calculation unit 6 is input to the expected loss calculation unit 7, and the expected loss value of each action is derived from this posterior probability value and the loss data input from the loss data processing unit 2. Execute the process. Then, these expected loss values are input to the action determining unit 8, and the action determining unit 8 executes a process of extracting one action having the smallest expected loss value. As described above, one appropriate action can be taken out.

A method of calculating the degree of similarity between two propositions in the similarity calculator 3 will be described. As a method for performing such a similar calculation, for example, Japanese Patent Application No. 4-315233 (Japanese Patent Application Laid-Open No. 6-1331) relating to the application of the applicant of the patent
The technical contents described in the specification of Japanese Patent Laid-Open No. 62099) can be adopted. This will be outlined below. This is a method in which an attribute list and an importance list are prepared for each proposition and the similarity calculation is performed using these lists. In this specification, the proposition is described as "horse". "Horse" attribute list is a list showing creatures, beasts, livestock, mane, etc. that are attributes of "horses", and an importance list is a creature, beast that is each attribute. A numerical value, for example, 0.8, 0.3, 0.6, 0.0, indicating the degree of importance of "horses" for livestock, mane, and others.
8 is a list indicated by. In this example, the degree of importance is displayed by a real value of 0.0 to 1.0.

By using this technical content, two
The similarity calculation between Proposition 1 and Proposition 2 is
Generally, it can be carried out as follows. First, of Proposition 1
Common included in the attribute list and the attribute list of Proposition 2
Focus on attributes. And about this common attribute, Proposition 1
Of the importance list of Proposition 2 and the importance list of Proposition 2
Similar to the value obtained by multiplying the value multiplied by the degree and the value added for all common attributes.
Degree. For example, the attribute list of Proposition 1 is (T_Two,
T_Three, T₈), The importance list is (0.3, 0.8, 0.
7) and the attribute list of Proposition 2 is (T_Three, T
₇, T₉), The importance list is (0.6, 0.5, 0.
2). At this time, the common attribute T _ThreeImportant to
The product of degrees is "0. 8 × 0. 6 = 0.48 ”, and
There is no common attribute, so common attribute T_ThreeThe product of the importance of
It becomes the value of similarity. In this way, Proposition 1 and Proposition 2 are
If they are similar, the number of these common attributes increases, and
The importance also increases with each other, so the product also increases, and
Therefore, the similarity, which is the sum of these products, takes a large value.
And On the contrary, if they are not similar, these common attributes
Since the numbers are smaller and their importance is smaller,
The similarity will take a small value. That is, similarity is a proposition
It shows the degree to which 1 and Proposition 2 are similar.

Further, the similarity calculation unit 3 can be configured to calculate the viewpoint similarity, which is the similarity between the two propositions when the "behavior" different from the previous attribute is viewed from the viewpoint (details are described below. , Japanese Unexamined Patent Publication No. 6-162099). For example, from the optimum action group (rain, umbrella) to the action group (fine,
When calculating the loss value of (umbrella), it is based on the premise of "taking the action of holding an umbrella" when {clear} and when {rain}
At that time, we attach importance to how different influences humans have. That is, the action {umbrella} is handled from the above viewpoint,
By providing the similarity calculation unit 3 with a configuration for calculating the similarity between {clear} and {rain}, the similarity between {clear} and {rain} corresponding to the premise of action {umbrella} can be obtained. This is expected to improve the accuracy of the similarity value. In this way, the similarity between two propositions when one action is adopted as a requirement for similarity calculation is called the viewpoint similarity. In order to configure the similarity calculation unit 3 that calculates the viewpoint similarity, a method based on the attribute list and the importance list can be adopted, as described above. In this case, the action is also regarded as one proposition, and an attribute list and an importance list are prepared for each action.
Here, when the viewpoint is not considered, the importance list is directly used for similarity calculation to calculate the similarity, but when considering the viewpoint, the importance lists of the two propositions are converted into the importance list considering the viewpoint. After conversion, the viewpoint similarity is calculated. In order to carry out the conversion of the importance list, the attribute included in the action attribute list among the attributes in the propositional attribute list has a large importance, for example, twice, and the attribute not included in the attribute list has a small importance. For example, it is set to 1/2. With this conversion,
The meaning of the action is reflected in each proposition. The propositional importance list thus obtained is called a modulation importance list. By using this modulation importance list, it is possible to obtain the viewpoint similarity reflecting the action as the viewpoint by calculating the similarity by the method focusing on the common attribute as described above. As described above, the similarity calculation unit 3 is configured to calculate the viewpoint similarity, and the value is used as the similarity. When the similarity calculation unit 3 that reflects the viewpoint is configured in this way, the loss data processing unit 2 is configured to pass the proposition set to the similarity calculation unit 3 as well as the action.

The loss calculation section 4 is configured to calculate a small loss value if the degree of similarity is large and a large loss value if the degree of similarity is small. The formula for calculating the loss value is
Although various types are conceivable, for example, there is a method of calculating as a value obtained by subtracting the degree of similarity from 1. The loss value calculation formula may be any formula as long as the similarity of the propositional pair is large, a small loss value is calculated, and if it is small, a large loss value is calculated. As described above, the generation and storage of loss data is completed.

The probability calculator 6 is configured to acquire the posterior probability value of each proposition from the facts and the probability data. For example, assuming the data of the format shown in FIG. 3 as the probability data, the posterior probability value of each proposition is acquired from this probability data.
The probability data is not limited to the format shown in FIG. 3, and any format of data can be adopted as long as the data can derive the posterior probability value. The probability calculator 6 is configured to calculate the posterior probability value of each proposition from such probability data based on an existing probability calculation formula.

The expected loss calculation unit 7 calculates the expected loss value of each action from the posterior probability value of each proposition supplied from the probability calculation unit 6 and the loss data supplied from the loss data processing unit 2. _{{W 1, ..., W n} } the set of all propositions When, the expected loss values of actions A when the facts F is given, "a posteriori probability value proposition W _i and action sets (Proposition W _i, The product of the loss values of the action A) is calculated as the sum of all propositions W ₁ , ..., W _n . The expected loss calculation unit 7 is configured to calculate the expected loss value for all actions as described above.

The action determining unit 8 is configured to select and present one action showing the minimum expected loss value out of the expected loss values of each action obtained by the expected loss calculating unit 7. Hereinafter, the procedure of action determination according to the present invention will be described more specifically with reference to FIG. In this case, the probability data shown in FIG. 3 is used, and the optimum action set shown in FIG. 5 is used. Then, the loss calculation unit 4 uses a method of calculating the loss value as “a value obtained by subtracting the degree of similarity from 1”.
The similarity calculation unit 3 uses the method of calculating the similarity between two propositions, and the similarity of (fine, rain), (fine, snow), (rain, snow) is 0.2, It is calculated as 0.1 and 0.8. However, the similarity of the same proposition is 1.0.

First, the loss data processing unit 2 takes out (clear, empty-handed) as an optimum action set. Then, the respective loss values of all the action groups (fine, empty-handed), (rain, empty-handed), (snow, empty-handed) including the extracted action “empty” of the optimum action set are calculated. When calculating the loss value of the action set (fine, empty-handed), the proposition “fine” of this action set and the proposition “fine” of the optimal action set (fine, clear) are passed to the similar calculation unit 3. The similarity calculator 3 calculates the similarity of this proposition set (fine, fine) as 1.0 and returns it to the loss data processor 2. The loss data processing unit 2 uses the calculated similarity of 1.0 as the loss calculation unit 4
Pass to. The loss calculation unit 4 returns the loss value 0.0 from the similarity 1.0 to the loss data processing unit 2. The loss data processing unit 2 acquires this loss value 0.0 as the loss value of the action group (fine, empty-handed).

In calculating the loss value of the action group (rain, empty-handed), the proposition "rain" of this action group and the proposition "fine" of the optimum action group
The proposition set (rain, fine) is passed to the similarity calculation unit 3. The similarity calculator 3 calculates the similarity of this propositional set (rain, fine) as 0.2,
It is returned to the loss data processing unit 2. The loss data processing unit 2 passes this similarity of 0.2 to the loss calculation unit 4. The loss calculation unit 4 returns a loss value of 0.8 to the loss data processing unit 2 based on this similarity. The loss data processing unit 2 acquires this loss value of 0.8 as the loss value of the action group (rain, empty-handed).

In calculating the loss value of the action set (snow, empty-handed), the proposition "snow" of this action set and the proposition "fine" of the optimum action set.
The propositional group (snow, fine) is passed to the similarity calculation unit 3. The similarity calculator 3 calculates the similarity of this propositional group (snow, clear) as 0.1,
It is returned to the loss data processing unit 2. The loss data processing unit 2 passes this similarity 0.1 to the loss calculation unit 4. The loss calculation unit 4 returns the loss value 0.9 from the similarity 0.1 to the loss data processing unit 2. The loss data processing unit 2 acquires this loss value of 0.9 as the loss value of the action group (snow, empty-handed).

Next, the loss data processing unit 2 takes out (rain, umbrella) as the optimum action set. And all the action groups (fine, umbrella) including the action "umbrella" of this optimal action group,
Calculate the respective loss values for (rain, umbrella) and (snow, umbrella). The loss values of these action sets can be calculated in the same manner as the calculation process in the case of the optimum action set (fine, empty-handed), and 0.8, 0.0, and 0.2 are obtained, respectively.

Further, the loss data processing unit 2 takes out (snow, umbrella and boots) as the optimum action set. Then, the loss values of all the action groups (fine, umbrella and boots), (rain, umbrella and boots), (snow, umbrella and boots) including the action “umbrella and boots” of the optimum action group are calculated. . The loss values for these behavioral groups can be calculated as above,
It is possible to obtain 9, 0.2, 0.0. As described above, the loss data can be obtained. The loss data thus obtained is summarized in FIG.

Now, the process for determining the action is started. First, the case where the state of the weather meter is {weather meter: high}, which is a fact, will be described. The probability calculation unit 6 uses the probability data shown in FIG. 3 for each weather “fine”, “rain”, and “snow”. The posterior probability values of 0.7, 0.2, 0.1, respectively.
To get. The expected loss calculation unit 7 calculates the expected loss value of the loss value for each action from the loss data obtained by the loss data processing unit 2 and the posterior probability data obtained by the probability calculation unit 6. The expected loss value when taking an action {umbrella} is L
(Clear, umbrella) x P (clear | high) + L (rain, umbrella) x P (rain |
High) + L (snow, umbrella) x P (snow | high) = 0.8 x 0.7+
It is calculated as 0.0 × 0.2 + 0.2 × 0.1 = 0.58. Similarly, the expected loss values for each of the actions “hand-held” and “umbrella and boots” are calculated as 0.25 and 0.67, respectively. The action determining unit 8 selects the action “hand-held” that takes the minimum expected loss value of 0.25. By presenting this "hands-free", it is possible to achieve the purpose of presenting an appropriate action for the questioner.

A case where the fact is {weather meter: medium} will be described. Probability calculation unit 6 is based on FIG.
The posterior probability values of “rain” and “snow” are 0.4,
Get 0.4 and 0.2. The expected loss calculation unit 7 is the loss data obtained by the loss data processing unit 2 and the probability calculation unit 6
An expected loss value is calculated for each action based on the posterior probability data obtained by. Expected loss value when taking action {umbrella} is L (fine, umbrella) x P (fine | medium) + L (rain,
Umbrella) x P (rain | medium) + L (snow, umbrella) x P (snow | medium) =
It can be calculated as 0.36. Similarly, the expected loss values for each of the actions “hand-held” and “umbrella and boots” are 0.5 and
Calculated as 0.44. The action determining unit 8 selects the action “umbrella” having the minimum expected loss value of 0.36. By presenting this "umbrella", the purpose of presenting an appropriate action for the questioner can be achieved.

A case where the fact is {weather gauge: low} will be described. Probability calculation unit 6 is based on FIG.
The posterior probability values of "rain" and "snow" are 0.1,
Get 0.4 and 0.5. The expected loss calculation unit 7 calculates an expected loss value for each action based on the loss data obtained by the loss data processing unit 2 "and the posterior probability data obtained by the probability calculation unit 6. The action {umbrella} is taken. Expected loss value is L (fine, umbrella) × P (fine | low) + L (rain,
Umbrella) x P (rain | low) + L (snow, umbrella) x P (snow | low) =
Calculated as 0.18. Similarly, the expected loss value for each of the actions “hand-held” and “umbrella and boots” is 0.7.
It is calculated to be 7, 0.17. The action determining unit 8 selects the action “umbrella and boots” having the smallest expected loss value of 0.17.
By presenting this “umbrella and boots”, it is possible to achieve the purpose of presenting an appropriate action for the questioner.

[0032]

As described above, when the loss data is acquired, a method in which a person directly writes a numerical value has been conventionally adopted. However, this loss data acquisition method causes various troubles to the human. Is a difficult method to obtain enough lost data to be able to answer the hesitation, and this difficulty has limited the problem that the reasoning device can handle to a certain narrow area. However, according to the present invention, even if the human being does not directly describe the loss value, the human being can directly describe the loss value by determining the action by using the loss value automatically calculated based on the similarity. You can achieve a decision equivalent to what you did. By presenting the action selected in this way, the purpose of presenting an appropriate action for the questioner can be achieved.

[Brief description of drawings]

FIG. 1 is a diagram illustrating an example.

FIG. 2 is a conceptual diagram of an inference device.

FIG. 3 is a diagram showing probability data.

FIG. 4 is a diagram showing loss data.

FIG. 5 is a diagram showing an optimum action group.

FIG. 6 is a diagram showing loss data calculated from similarity.

[Explanation of symbols]

 2 Loss data processing unit 3 Similarity calculation unit 4 Loss calculation unit 6 Probability calculation unit 7 Expected loss calculation unit 8 Behavior decision unit

Claims (4)

[Claims] 1. A plurality of propositions whose values are probability values,
Actions in an inference device that selects a certain action by using certain facts, probability data that derives the posterior probability value of the proposition based on the facts, and the optimal action set of action sets that combine one action with the proposition In the determination method, the proposition of any two action sets is associated with a similarity consisting of real numbers of 0.0 or more and 1.0 or less, and any one action set is associated with a real number of 0.0 or more and 1.0 or less. When associating with a loss value consisting of numerical values, the optimal action set is associated with 0.0, and the proposition and similarity of the optimal action set are s.
With regard to the action set of the proposition that is, the small loss is associated with the small value when s is large, and the high loss is associated with the large value when s is small. A method for deciding an action in an inference device, which is characterized by deriving a value and selecting one action having the smallest expected loss value. 2. The action determining method in the inference device according to claim 1, wherein the similarity is a viewpoint similarity which is a similarity between two propositions when an action is viewed from a viewpoint. Behavior decision method. 3. An action set having a loss data processing unit to which an optimum action set (proposition W, action A) and an action set (proposition X, action A) are input, and having the action A passed from the loss data processing unit. The propositional set of the propositional set passed from the lossy data processing unit is provided with a similarity calculation unit that calculates the similarity of the propositional set (proposition W, propositional X) consisting of the proposition X and the proposition W of Probability of calculating the posterior probability value of the proposition based on the given facts and probability data, with the loss calculator that calculates the loss value of the action group based on the similarity and returns the calculation result to the loss data processor. It is equipped with a calculation unit, and has an expected loss calculation unit that calculates expected loss data based on the loss data and posterior probability values, which are the set of loss values of the action set returned to the loss data processing unit, and is input. Expected loss data Action decision device in the inference apparatus characterized by comprising the action determining unit to retrieve the minimum expected loss values. 4. The action determination device in the inference device according to claim 3, wherein the similarity in the similarity calculation unit is converted into the viewpoint similarity, which is the similarity between two propositions when the behavior is viewed from the viewpoint. An action determining device in an inference device characterized by being.