JP4432547B2 - Inference system - Google Patents

Description

The present invention relates to a reasoning system for executing inference by using the rules expressed in If-Then format.

  Conventionally, there is known a rule-based reasoning system that expresses expert knowledge in a rule format (If-Then format) and finds a solution by applying a rule to a given problem (for example, non-patented) Reference 1).

  Here, an example of rules used in this kind of inference system is shown in (1). This rule is used when determining a destination suitable for the situation description based on the given situation description.

IF {(season = summer) AND (weather = clear) AND (hobby = swimming)} THEN {sea} (1)
In other words, the conditional clause of the rule contains linguistic or numerical conditions (here “summer”, “sunny”, “swimming”) for all attributes to be considered (here “season” “weather” “hobby”). A logical judgment is described for the status description of satisfaction or not, and the subordinate clause of the rule is one judgment that the system should make if the conditions indicated in the conditional clause are satisfied (here, “sea”) Is described.

  By the way, as one of the applications of such an inference system, according to the inference result obtained from the inference system (contents of subordinate clauses), the control corresponding to the inference result (for example, “the system sends an emergency message to the center” "Make music", "Turn on TV and show news", "Bake a bath", etc.).

  In the inference system, depending on how the rules are created, there may be a plurality of rules (that is, a plurality of inference results can be obtained) that hit (satisfy a conditional clause) for a given situation. In such a case, for example, the above-described application device executes the control only for the first hit rule that satisfies the conditional clause, or executes the control for all the hit rules in the order of the hit. Can be considered.

  However, in this case, a less valid rule may be followed by a more valid rule, in which case the inference system made a decision that was not necessarily optimal (or possibly wrong). It will be. This is equivalent to having fallen into local optimization, for example, in the optimal value search problem.

In order to avoid such a problem, it is necessary to configure all rule groups so that only one rule is hit for one situation description.
Masashi Yoji, Makino Michika, Saedaburo Saburo, Tarumi Hiroyuki “Mobile Field Service Workflow System Emphasizing Strategic Decisions” Information Processing Society 43rd Groupware and Network Services Study Group, GN-43-20 (2002)

  However, in order to do so, the logic of the conditional clause of each rule must be completely contradictory (equivalent to dividing the space of the situation description into direct sums), and the situation is described in a completely divided hierarchy. Is required.

FIG. 14 shows the hierarchical rules in a table, where there are N status description attributes, and the resolution of each attribute i (i = 1 to N) is M _i. The number of situations obtained by the combination, that is, the number of rules Ns is expressed by (2).

Ns = M ₁ × M ₂ × ... × M _N (2)
The number N of attributes increases as various condition descriptions such as the user's environment, situation, request, and status are reflected in the conditional clause, and the determination accuracy for each attribute increases. The resolution Mi increases. As a result, the number of rules to be described becomes enormous, and in some cases, there is a problem that the level becomes unrealizable.

  In this case, as can be seen from the equation (2), when the attribute of the situation description is added later (N is increased), a huge number of rules must be added accordingly, There was also a problem that the addition could not be performed easily.

  Furthermore, when all the combination conditions considered in this way are ruled out and described one by one, the condition description described in the conditional clause does not have a remarkable feature or cannot actually occur. There are also problems that describing rules in this way is inefficient.

  Note that it is possible to reduce the number of rules by implementing only rules that are considered important for application, but in that case, not only will it not be possible to make detailed inferences, but also which rules (attributes) are important. There was also a problem that it was difficult to judge.

In order to solve the above-described problems, an object of the present invention is to provide an inference system in which rules can be easily added and changed, and adaptation according to an individual and a situation can be dynamically performed.

The inference system of the present invention made to achieve the above object uses items set to classify various situations as situation attributes, and any one of a plurality of situation attributes set in advance as a conditional clause. A linguistic or numerical condition for one or more is described, and an item representing a driveable application or an estimated user's desire is set as a request attribute and a subordinate clause. A rule storage unit is provided that stores a rule group including a large number of rules in which scoring vectors whose components are required degrees are described.

Then, the scoring vector extraction means acquires the situation data indicating the user situation and the environment surrounding the user, and based on the situation data, the satisfaction of the condition clause of the rule stored in the rule storage means is satisfied or not satisfied Is determined based on whether or not the description of the situation attribute constituting the conditional clause matches the situation data, and for all rules determined to satisfy the conditional clause, the score constituting the dependent clause of the rule Extract the ring vector.

Then, the request vector generation unit performs a cumulative process of weighting and adding the scoring vector extracted by the scoring vector extraction unit using a weighting factor set in advance to represent the importance of the scoring vector. Then, a request vector having a request level for each of the request attributes in the situation comprehensively grasped from the situation data as a component is generated, and the inference result generating means generates an inference result based on the generated request vector. .

  In other words, a conditional clause is configured to allow multiple rules to be hit (a conditional clause is satisfied) for a certain situation setting, and the scoring vector constituting the subordinate clause is set for all the hit rules. By superimposing, a request vector that reflects all the hit rules is obtained, and an inference result is obtained according to this request vector.

Therefore, according to the inference system of the present invention, even if a plurality of rules are hit, it is possible to obtain an optimal inference result overlooking the hit rules without falling into local optimization.
Further, in the inference system of the present invention, since the situation description is not performed for all attributes at once, but is divided into a plurality of rules, the group of rules having different situation attributes (or combinations thereof) used for the conditional clauses. Each time, an independent inference process can be executed.

  In addition, since the description of the conditional clause is independent for each rule group, not only can the number of rules to be set be significantly reduced compared to the conventional device including all the status attributes in the conditional clause, It makes it easy to add and modify rules, and thus dynamically adapt to individuals and situations.

  Further, in the inference system of the present invention, by describing the subordinate clause of a rule using a scoring vector, it is possible to express the characteristic of the determination result while leaving ambiguity in the subordinate clause in each rule.

  By the way, the rule group stored in the rule storage means may be composed of only one-dimensional rules in which linguistic or numerical conditions for a single situation attribute are described in the conditional clause, or a plurality of conditional clauses may be included in the conditional clause. It may be configured only from a plurality of hierarchical rules in which linguistic or numerical conditions for the situation attribute are described, or may be configured to include both one-dimensional rules and hierarchical rules.

Since the one-dimensional rule includes only one situation attribute in the conditional clause, the degree of contribution of the situation attribute to the required attribute is clear, and rule generation and correction can be easily performed.
On the other hand, the hierarchical rule is suitable for a description when a specific logical condition represented by a plurality of situation attributes has a special meaning, for example, a description about a special case or an exception process.

  However, in a hierarchical rule, the more situation attributes included in a conditional clause, the more difficult it is to generate and modify the rule. Therefore, priority is given to rules that do not require much addition, modification, or modification, and simplification of processing. It is desirable to apply to a rule to be used, and to use a situation attribute having a relatively small resolution by reducing the number of situation attributes constituting a conditional clause.

  In the inference system of the present invention, a plurality of rule groups may be stored in the rule storage unit, and the scoring vector extraction unit may be configured to select a rule group to be used according to the situation data. In this case, since only a rule group according to the situation is used in a limited manner, the inference process can be executed adaptively and efficiently (and thus at high speed).

  The situation data includes information obtained from user profiles (including personal information, user type, natural environment, information environment, situation, request, and state), time changes, and place changes (including country changes). The situation limited information may be used. In addition, as the rule group according to the situation, for example, a rule group classified based on geographical conditions (for example, a rule group for Japan / US / China), a rule classified based on a user type Groups (for men / women / children / adults, etc.), rules grouped according to purpose (rules for work / excursions, etc.), etc. Other functions (subordinate clauses) ) And status groupings (corresponding to conditional clauses) may be used.

When a driveable application is associated with a request attribute, an application to be driven can be immediately inferred from a request vector obtained by accumulating scoring vectors.
Further, when the estimated user's desire is associated with the request attribute, the user's desire is, for example, “I want to eat”, “I want to play sports”, “I want to relax”, “I want to change my mood”, “I want to shop”, “I want to be fashionable” “I want to talk to someone”, “I want to go somewhere”, “I want to watch a movie”, “I want to listen to music”, “I want to play a game”, etc. What is represented by the potential state may be used.

In the present invention, the request vector generation means includes attribute conversion means for converting a request vector having a user attribute obtained by accumulation processing as a request attribute into a request vector having a driveable application as a request attribute. Is desirable.

  If such an attribute conversion means is provided, it is not necessary to change the scoring vector at the time of adding a new application, and the association with the added application can be performed simply by changing the attribute conversion means. be able to.

Further, a scoring vector is set to represent the average user characteristics , and the situation data includes a deviation of individual user characteristics relative to the average user characteristics, a coordinate transformation matrix and a translation vector. If the viewpoint data is included at least expressed in the request vector generation means, a request vector obtained by the cumulative, by conversion using the viewpoint data, generates a request vector that reflects the characteristics of the user It is desirable to provide viewpoint conversion means.

  In this case, the rule group (especially the scoring vector) itself does not need to reflect the characteristics (viewpoint) of each user, and the same rule group can be applied to all users. An increase in scale can be suppressed.

Incidentally, scoring vector is equivalent to what is represented using a common coordinate system corresponding to the average user's perspective, since the realized viewpoint conversion by using the technique of such coordinate transformation is there.

As the initial data of the scoring vector, for example, data obtained from the result of statistically averaging the evaluations of attribute values based on subjective judgments of a plurality of users can be used. In addition, as a totaling method of evaluation, a questionnaire (the use of the Internet is also considered), a probe system, etc. may be used.

  In this case, it takes a lot of time and effort to collect data, and fragmentary rules of thumb that can be easily inferred from common sense (for example, “Traffic jams occur in Bon Festival, winter holidays, and Golden Week in Japan. It is relatively difficult to explicitly reflect "news needs".

Therefore, as the initial data of the scoring vector, for example, a subjective empirical rule of a specific user (for example, a system developer) (especially, a typical aspect corresponding to the above-described example of the fragmentary empirical rule is focused on) It is also possible to use a digitized version of

  In this case, since it is difficult to define a rule group corresponding to all the situations and situations that many users face, based on this small number of representative rule groups, rules for unfilled places (especially scoring vectors) It is desirable to use a method that infers).

The initial data of the scoring vector may be set in advance, but may be configured to be set or reset by the initial data setting unit.
The initial data setting means may set, for example, externally input data or a template selected by an external operation from a preset template group as initial data.

  Even if the initial data is set by the above-described method or a combination of these methods, the amount of work in the initial data setting means becomes enormous, so the score for all possible situations (rules) It is difficult to prepare initial data of a ring vector.

Therefore, the initial data setting means uses a part of the situation attribute included in the rule as a partial situation attribute, and sets the rule group as a plurality of partial rules for each rule in which a condition related to the same partial situation attribute is described in the conditional clause. An initial data transfer means for transferring the same initial data to the scoring vector of another rule belonging to the same partial rule group as the rule when the initial data of the scoring vector is set for a certain rule. It is desirable to provide. According to such initial data transfer means, it is only necessary to set one initial data for each partial rule group, so that the rules to be set by the initial data setting means can be greatly reduced.

Also, the inference system of the present invention uses the request vector used when the inference result generation means generates the inference result and the response data indicating the response of the user who received the inference result as feedback information for the inference result. , A threshold value set in advance between an operation vector generated from the response data and representing the user's actual request in the same format as the request vector, and a request vector associated with the response data by feedback information. If there is a difference greater than the value, or if it is statistically recognized that there is a difference greater than the threshold value between the two vectors from the history of feedback information within a preset period, it is assumed that the correction condition is satisfied. The inference characteristic correction means is based on a target rule that is a rule used to generate a request vector that satisfies the correction condition. Te, so that the difference between the operation vector request vector becomes smaller, by changing the value of scoring vectors constituting the subordinate clauses of target rules, it may be configured to modify the inferences characteristics of the system .

In this case, the feedback information used for correcting the inference characteristic may relate to the same user or may relate to another user . Further, the history of feedback information may be stored in the history storage means.

In the inference system of the present invention, the response detection unit detects a response from the user who received the inference result generated by the inference result generation unit, and the detected response affirms the inference result. If it is, the rule strengthening means may be configured so that the rule related to the request vector used for obtaining the inference result is the rule to be strengthened and the rule to be strengthened is strengthened. Here, even when there is no “response that denies the inference result”, it is included in the “response that affirms the inference result”.
However, the rule strengthening means changes the weighting factor used when generating the request vector from the scoring vector constituting the dependent clause of the rule to be strengthened by the request vector generating means, or the dependent clause of the rule to be strengthened The rule to be strengthened is strengthened by increasing the value of the scoring vector that constitutes.

The rules reinforcing means, for conditional enhancements include rule, part of the status attribute included in the enhanced receiver rule as part status attribute, rule condition relating to the same parts status attribute to the condition clause is described (hereinafter (Also referred to as “neighboring rules”) may be targeted for strengthening.

However, when strengthening a nearby rule, it is desirable to suppress the strengthened part from the original rule to be strengthened .
As described above, according to the inference system of the present invention, it is possible to modify or enhance the inference characteristics, and thus it is possible to dynamically perform adaptation according to the individual and the situation.

Embodiments of the present invention will be described below with reference to the drawings.
First, FIG. 1 is a block diagram showing the configuration of an interface device 1 to which the present invention is applied. The interface device 1 of the present embodiment is mounted on an automobile (vehicle), and various applications (for example, various devices using various devices mounted on the vehicle) are used for a vehicle occupant (mainly a driver). Providing video display, music playback, Web search, server connection, voice recognition, etc.) with a simple interface.

  As shown in FIG. 1, the interface device 1 according to the present embodiment includes an input including a selection key 41 to 43 used for selecting an application and an operation key group 44 for inputting various commands and data by an external operation. Display provided with device 3, assigned application display sections 51 to 53 for displaying application names associated with the selection keys 41 to 43, and provided application display sections 54 for displaying the names of applications being provided The apparatus 5 is connected to a display 7 for displaying an image and a microphone 8 for inputting sound.

  In the operation key group 44, a status setting used for inputting a start key, a correction key, and status data (for example, season, weather, destination, destination, information, user status, etc.) used for processing to be described later. At least the key is included.

  In addition, the selection keys 41 to 43 are formed of push button switches formed integrally with the assigned application display units 51 to 53, respectively, and are arranged side by side at the lower portion of the display 7, so that the vehicle occupants (especially drivers) can easily recognize them. In addition, a shape, color, and size that are easy to operate are used. Further, the operation key group 44 and the provided application display unit 54 are arranged in the vicinity of the selection keys 41 to 43.

  The interface device 1 includes a well-known navigation device 11 for detecting the current position (current location) of the vehicle and route guidance, an air conditioner device 13 for controlling air conditioning in the vehicle, a cassette tape recorder, and a CD (compact disc) player. , An MD (mini-disc) player, a radio, a television, and the like, and a wireless terminal between a well-known VICS (Vehicle Information and Communication System) broadcasting terminal and an Internet broadcasting terminal serving as a connection window with the Internet. , A communication device 17 for performing data communication, various sensors 19 for detecting the vehicle operating state such as the vehicle speed and acceleration / deceleration state, the temperature inside and outside the vehicle, the presence or absence of raindrops, etc. Window), engine, and other control devices for controlling brake devices ( It is connected to indicate shown) and.

  The navigation device 11 is a known GPS device for detecting the current position of the vehicle, a CD-ROM that stores route guidance data such as map data, place name data, and facility name data, and data from the CD-ROM. A CD-ROM drive for reading out and an operation key for a user to input a command. The navigation device 11 is optimal for reaching the current position of the vehicle and the destination when, for example, an instruction for instructing the route to the destination and the destination is input via the operation key from the user. A road map including a simple route is displayed on the display 7 to provide route guidance.

  The display 7 not only displays a road map for route guidance by the navigation device 11, but also displays various images such as an information search menu, and the audio device 15 is set to the TV mode. Then, the received image of the television received by the television tuner provided in the audio device 15 is displayed.

  The interface device 1 includes a system control unit 21 mainly composed of a microcomputer including a CPU, a ROM, a RAM, and the like, and an input interface (I) for inputting commands and data from the input device 3 to the system control unit 21. / F) unit 23, display control unit 25 that controls the display of each unit 51 to 54 constituting display device 5, screen control unit 27 that controls the display screen of display 7, and audio signal input from microphone 8 A voice input unit 28 that converts the data into digital data and inputs the digital data to the system control unit 21, the display 7, the navigation device 11, the air conditioner device 13, the audio device 15, the communication device 17, various sensors 19, and other control devices. A device control interface (device control I) that connects the system control unit 21 to enable data communication. And a F) 29.

  In addition, the interface device 1 is provided with an Internet address database 31 for storing an Internet address (Internet address) and a search control unit 33 in order to search and acquire desired information from the Internet by the communication device 17. Yes. Then, when the system control unit 21 outputs a search keyword representing the search content to the search control unit 33, the search control unit 33 operates the communication device 17 via the device control I / F 29 to perform the search from the Internet broadcast terminal. Information corresponding to the keyword is searched, and the search result is input to the system control unit 21. The Internet address database 31 stores Internet addresses used in the past by the search control unit 33 according to commands from the system control unit 21, and the search control unit 33 searches the system control unit 21 in the past. When the same search keyword as the keyword is received, the Internet address in the Internet address database 31 is reused.

  Further, the interface device 1 includes a rule base 35a describing a conditional clause of an inference rule in If-Then format, a scoring vector table (SVT) 35b describing a subordinate clause of the inference rule, and an inference using the above inference rule. A rule storage unit 35 for storing a situation storage table (SMT) 35c in which results and user operation histories are described in association with each other, and situation data (for example, a natural environment, an information environment) indicating a situation such as an environment surrounding the user or the user , Interest information, user status, etc.) and a situation data storage unit 37 for storing a user profile 37b describing personal information of a plurality of users.

  Among these, the situation data described in the situation data file 37a may use data other than those exemplified as long as it affects the user's request and action. In addition, “natural environment” listed as an example of situation data refers to seasons and weather, and “information environment” refers to information on terminals and networks for acquiring various types of information (for example, terminal model, Terminal capability, network capability, etc.). The “interest information” is information that is important to many people (for example, “Typhoon is approaching”, “World Cup is being held”, etc.). The user situation indicates the user's behavior (the purpose or destination of the behavior), physical condition, mood, etc. at that time.

  The personal information described in the user profile 37b stores individual user preferences in association with various situations such as time situations (time zones, seasons, anniversaries, etc.), geographical situations, and mood situations. Is.

  As shown in FIG. 2, the rule base 35a is a table in which rule numbers are associated with each attribute value of a preset situation attribute or each combination of attribute values. However, inference rules include a one-dimensional rule in which only one situation attribute is described in the conditional clause, and a hierarchical rule in which a plurality of situation attributes are described. In FIG. Is shown in (a), and a description of the hierarchical rules is shown in (b).

  The situation attributes include, for example, “season”, “day of the week”, “time zone”, “purpose”, “destination”, “weather”, etc., and these situation attributes are linguistic or numerical conditions for the situation attributes, respectively. (For example, if the status attribute is “season”, the attribute value is “spring”, “summer”, “autumn”, “winter”).

  Further, in the rule base 35a, situation attributes whose importance does not change greatly even when combined with other things are described mainly by one-dimensional rules, and situation attributes whose importance changes greatly depending on combinations with other things ( For example, those that describe special cases and exception handling) are mainly described by hierarchical rules. In addition, the rule base 35a is configured to ensure independence between individual rules regardless of one-dimensional rules and hierarchical rules.

Further, the rule base 35a stores a plurality of rule groups grouped for each function (corresponding to a subordinate clause) and situation classification (corresponding to a conditional clause).
Specific examples include "rules for Japan / USA / China", "rules for men / women", "rules for children / adults", and "rules for work / excursions". Each rule group may be configured to share some rules as shown in the conceptual diagram of FIG.

  As shown in FIG. 3A, the SVT 35b, for each rule number Ri defined by the rule base 35a, has a scoring vector vi having as a component a degree of request for each of the preset request attributes, and its scoring. It is a table that defines a weighting factor wi that represents the importance of a vector vi.

  In addition, as a request attribute, what expresses the user's desire is used, for example, "I want to eat" "I want to play sports" "I want to relax" "I want to change my mood" "I want to shop" "I want to be fashionable" "I want to talk to someone "I want to go somewhere", "I want to watch a movie", "I want to listen to music", "I want to play a game", etc.

However, in the present embodiment, the attribute value (that is, the request level) of the request attribute is expressed by a numerical value of 0 to 100, and the weighting coefficient is expressed by a numerical value of 0 to 1.
Further, initial data is given in advance to each scoring vector described in the SVT 35b, and the contents thereof are set based on average user characteristics ( viewpoint ) .

  In this embodiment, for the scoring vector of the one-dimensional rule, specifically, attribute values based on subjective judgments of many users are used by using a questionnaire (considering use of the Internet) or a probe system. The result of totaling and statistically averaging the totaled data is set as initial data.

  In addition, for the scoring vector of hierarchical rules, a subjective rule of thumb (representative of situations where a combination of specific conditions has great meaning) by a small number of people (system developers, etc.) is quantified. Is set as initial data.

  However, only a collection of subjective heuristics by a small number of humans should specify all scoring vectors (corresponding to a group of rules describing all aspects or situations that many users face) that constitute SVT 35b. Therefore, based on this small number of representative subjective rules of thumb, unfilled portions of the SVT 35b are filled by using a method such as inference.

  The SVT 35b may include only one that is commonly used for all rule groups constituting the rule base 35a. However, when the rule groups are grouped by function, each rule group (That is, a plurality of SVTs may be provided). In this case, the type and number of request attributes (scoring vector dimensions) corresponding to each component of the scoring vector are also different for each rule group. May be different.

Status storage table (SMT) 35c, as shown in FIG. 3 (b), as a result of the inference processing using the above inference rules, request vector (Application vector) that is provided to select the key setting process to be described later and, Feedback information from the user regarding the setting of the selection keys 41 to 43 based on the request vector (application selection operation, setting change operation of the selection keys 41 to 43, etc.) is stored in association with the date and time when the feedback information was obtained. It is a table to do.

  Note that the rule storage unit 35, the status data storage unit 37, and the above-described Internet address database 31 are configured by a nonvolatile memory capable of reading and writing data.

  Next, in the interface device 1 configured as described above, an outline of the main processing executed by the system control unit 21 will be described with reference to the functional block diagram shown in FIG. 4 and the flowcharts shown in FIGS. .

  In addition to the main process, the CPU configuring the system control unit 21 executes a situation data update process for updating the contents of the situation data file 37a based on information collected through the communication device 17 and various sensors 19.

  In addition to the program for executing these main processing and status data update processing, the ROM that constitutes the system control unit 21 performs processing and device group control necessary for application providing processing executed in the main processing. Various application programs (for example, a video display application, a music playback application, a Web search application, a server connection application, a voice recognition application, etc.) for execution are stored.

  As shown in FIG. 5, when the main process is started, it is first determined whether or not a start key that is one of the operation key groups 44 has been operated (S100). An inference process for estimating the request is executed (S110).

  In this inference process, as shown in FIG. 6, first, the situation data file 37a and the user profile 37b are read from the situation data storage unit 37 (S210), and based on the read contents, the rule base 35a of the rule storage unit 35 is read. A rule group to be used at the present time is selected from the list (S220). At this time, the total number of rules belonging to the selected rule group is m, and in the following, these rules are represented by R1 to Rm.

  Then, the parameter i used for identifying the rule is initialized to zero, and based on the SVT corresponding to the selected rule group, the same N-dimensional request vector as the scoring vector described in the SVT (the number of components is the same) V is set, and the value of each component of the request vector V is initialized to zero (S230).

  Thereafter, the parameter i is incremented (S240), and it is determined whether or not the condition described in the conditional clause of the rule Ri is established based on the current situation specified from the content read in S210 (S250).

  If the condition described in the conditional clause of rule Ri is satisfied, the scoring vector vi and weighting factor wi of rule Ri are read from SVT 35b (S260), and as shown in equation (3), the scoring vector vi is read. The request vector V is updated by adding the product of the weight coefficient wi to the request vector V (S270), and the process proceeds to S280.

V ← V + wi * Vi (3)
On the other hand, if it is determined in S250 that the condition described in the conditional clause of the rule Ri is not satisfied, the process proceeds to S280 without executing S260 and S270, and the parameter i is greater than or equal to the total number m of rules (i ≧ m) is determined (S280). If i ≧ m is not satisfied, it is determined that there is an unprocessed rule in the selected rule group, and the process returns to S240 to perform the above-described processing (S240 to S270). )repeat.

If i ≧ m, the above processing is completed for all rules belonging to the selected rule group, and the request attributes associated with each component of the request vector V are converted from user desires to applications. attribute conversion process is executed (S290), further, a request vector attribute conversion processing has been performed (application vector) Va, based on the contents of the user profile 37b read in previous S210, the individual user characteristics ( request vectors viewpoint) is reflected (by executing the viewpoint conversion processing to convert the application vector) Vp (S300), the process ends.

  That is, by repeating the processes of S240 to S270, the result of weighted addition for the scoring vectors of all the rules that are hit (conditions are satisfied) is the request vector V.

FIG. 8 is an explanatory diagram schematically showing this processing. However, the case where the selected rule group includes only one-dimensional rules related to the _M situation attributes A _{1 to} A _M is illustrated.
In addition, a one-dimensional rule in which only a condition related to the situation attribute A _i is described in the conditional clause is represented by (4).

Rule [A _i , C _i (J _i ), v _i (J _i )] (4)
C _i (J _i ): A conditional clause of the J _i th rule in the rule group relating to A _i v _i (J _i ): Scoring vector acquired when C _i (J _i ) is satisfied That is, each situation attribute A _i 0 or 1 rule is selected from each of the partial rule groups for N, and an N-dimensional scoring vector corresponding to the selected rule is assigned to each of the coordinate axes X _{1 to} X _N corresponding to the situation attributes. The request vector V is obtained by superimposing on the dimensional coordinate. In addition, FIG. 9 illustrates a case where “time zone”, “season”, and “day of the week” are used as situation attributes more specifically.

  Further, in the attribute conversion process of S290, each component (request attribute) is associated with the request level for the desires of N types of users, and each component (request attribute) is associated with L types of applications (contents). It is converted into a request vector (hereinafter also referred to as “application vector”) Va associated with the request level.

Examples of the request attribute of the application vector Va include “news”, “road information”, “weather information”, “closest facility information”, “map display”, “calm music”, and the like.
The application vector Va is specifically calculated by performing the vector operation shown in (5) on the request vector V obtained by the accumulation (weighted addition) process (see FIG. 10).

Va = PV + f (5)
P: L × N-dimensional conversion matrix for converting an N-dimensional vector into an L-dimensional vector f: an L-dimensional constant vector corresponding to the statistical average of the application vector Va Note that the conversion matrix P is, for example, a covariance structure analysis By using it, it can be determined from past statistical data. The covariance structure analysis is a well-known method and has no relation to the main part of the present invention, so that the description thereof is omitted here.

Further, in the viewpoint conversion process of S300, the application vector Va can be regarded as being expressed using the common coordinate system O corresponding to the average (neutral) viewpoint (feature) of the user. The viewpoint conversion is realized by performing coordinate conversion to a personal coordinate system corresponding to the personal viewpoint (feature) .

Specifically, the coordinate conversion corresponding to the individual viewpoint is realized by performing the vector operation shown in (6) using the L × L-dimensional coordinate conversion matrix R _A and the L-dimensional translation vector d _A. .
_{Vp = R A (Va-d} A) (6)
Note that the coordinate transformation matrix R _A and the translation vector d _A may be set in advance as data of the user profile 37b.

For example, when the coordinate transformation matrix _RA is a unit matrix I for simplicity, (6) is represented by (7).
Vp = Va-d _A (7)
Furthermore, the application vector Va is two-dimensional (L = 2), the request attributes are “news” and “jazz music playback”, the application vector Va is represented by (8), and the translation vector d _A is represented by (9). Think.

Va = (request for news, request for playing jazz music) ^t
= (100, 80) ^t (8)
d _A = (30, −10) ^t (9)
The application vector Vp after coordinate conversion calculated based on (7) to (9) is represented by (10).

Vp = (70, 90) ^t (10)
That is, even if it is estimated that the “news demand” in a certain situation is 100 for the average user, it is 70 for the user A (from the viewpoint of the user A). The more users are, the less demanding news is. On the other hand, the “requirement level for playing jazz music” is estimated to be 80 for the average user, but is 90 for the user A. Therefore, the average user and the user A have no restrictions on the application. The degree of demand will be reversed.

  Returning to FIG. 5, when the inference process (S110) is completed, the selection key setting process is executed based on the application vector Vp obtained as the inference result (S120), and the process returns to S100.

  In this selection key setting process, first, application names associated with each component of the application vector Vp are in descending order of component values, that is, in descending order of user demand (hereinafter referred to as “application recommended rank”). Generate a side-by-side application list. Then, the three applications at the top of the application list are associated with the selection keys 41 to 43, respectively, and the names of the associated applications or labels indicating the applications are displayed on the assigned application display units 51 to 53.

  That is, when the start key is operated, the best three applications that are estimated to be requested by the user are selected from the contents of the situation data file and the user profile, and assigned to the selection keys 41 to 43. In other words, the interface apparatus 1 recommends an application according to the situation to the user.

  Next, when it is determined in S100 that the start key is not operated, it is determined whether or not the selection keys 41 to 43 are operated (S130), and the selection keys 41 to 43 are operated. If so, an application providing process for providing an application associated with the operated selection key using various devices and sensors connected via the device control I / F 29 is executed (S140). The feedback information indicating the user's operation (response) that triggered the providing process and the inference result (the application vector Vp or Va or the request vector V) in the inference process of S110 are stored in the situation storage table 35c in association with each other. The history saving process is executed (S150), and the process returns to S100.

  In the application providing process of S140, not only the application is provided, but also a process of displaying the name of the application being provided or a label indicating the application on the provided application display unit 54 is executed. In addition, when a special operation (for example, long press) is applied to the selection key, processing for changing the association of the application to the operated selection key is executed according to the application list created in S120. .

  In the history saving process of S150, when an application is provided by operating the selection keys 41 to 43 in the application providing process, the provided application is used as feedback information, and the selection keys 41 to 43 are provided. When the application associated with is changed, the changed content is used as feedback information.

  The contents of the status storage table 35c (hereinafter referred to as “operation data”) stored by the history saving process are transmitted to an external information accumulation center via the communication device 17 by a separately started communication process. The

  Here, when any one of the selection keys 41 to 43 is operated, the application assigned to the operated selection key is driven. However, the application assignment to the selection keys 41 to 43 is performed. It may be configured to automatically provide the application with the highest application recommendation rank immediately after the operation is performed or when the selection keys 41 to 43 are not operated for a preset standby time or longer. Further, even when such an application is automatically provided, it is stored in the status storage table 35c as feedback information indicating that there is no negative response from the user (the user has made a positive response).

Then, if in the previous S130, it is determined that the selection key 41 to 43 has not been operated, in turn, preset adjustment conditions it is determined whether or not satisfied (S160), Osamu positive conditions Is established, an inference characteristic adjustment process for correcting the inference characteristic in the inference process of S110 is executed (S170), and the process returns to S100.

The correction conditions determined in S160 are as shown in the following (A) to (D).
(A) The application vector Va (that is, the estimated request) stored in the status storage table 35c in the history storage process of S150, and the operation vector (that is, the actual request) generated based on the feedback information The difference between them (size, angle, etc.) is greater than or equal to a preset threshold.

  (B) As a result of statistical determination based on operation data (application vector, feedback information) stored in the status storage table 35c over a certain period in the past, the condition shown in (A) is noticeable. .

(C) When the user inputs an inference characteristic correction request via the operation key group 44.
(D) When there is no negative reaction from the user (that is, an operation for changing the assigned application) to the setting by the selection key setting process.

  In the inference characteristic adjustment process of S170, when the correction condition (A) or (B) is satisfied, the application vector Va is generated so that the difference between the application vector and the operation vector becomes small. Modify the inference characteristics of the part involved.

However, in this case, specifically, the following (a) to (f) are to be corrected.
(A) Scoring vector value (b) Rule-based description (addition and subdivision of conditional clauses)
(C) Parameters used for accumulation processing (weighting coefficient wi)
(D) Transformation matrix P used for attribute transformation processing, constant vector f
(E) Coordinate transformation matrix R _A and translation vector d _A used for viewpoint transformation processing
(F) Description of user profile used for determination of conditional clause In addition, when the correction condition of (C) is satisfied, a group of operation data collected by another device through communication with the information accumulation center via the communication device 17 The new rule extracted by using the acquired operation data group or using the operation data group stored in the situation storage table 35c and performing statistical processing on the operation data group Are added to the rule base 35a and the SVT 35b.

  When the correction condition (D) is satisfied, the inference characteristics (particularly individual rules) of the part related to generation of the application vector Va provided for the selection key setting process are strengthened. In this case, the above-described (a) and (c) are to be strengthened. If the vicinity of the rule to be strengthened (center rule) can be defined, the rule belonging to this vicinity (neighbor rule) is also strengthened.

  However, for the neighborhood rule, the increment in the central rule multiplied by the coefficient value (0 to 1) is defined as the increment so that the increment in the strengthened value does not exceed the increment in the central rule. Have been to. Further, the neighborhood may be, for example, a rule having an attribute value belonging to the same category as the attribute value indicated in the central rule for the situation attribute described in the conditional clause.

  Next, when it is determined in S160 that the correction condition is not satisfied, it is determined whether or not an initial data setting request for the SVT 35b is input via the operation key group 44 (S180). ) If a setting request is input, the initial data setting process is executed (S190), and the process returns to S110.

In the initial data setting process in S190, one of the following (i) and (ii) is executed.
(I) Data input by operating the operation key group 44 is stored in the SVT 35b as initial data. However, the operation with the operation key group 44 may be to individually set individual data, or by selecting a preset template group from a plurality of preset data. May be set. In addition, it is desirable to use an interactive support tool prepared in advance for these inputs.

  (Ii) Data generated by performing inference processing based on the description content of the user profile 37b is stored in the SVT 35b as initial data. Note that, for example, this method is used in a system in which SVT is set for each user without performing the viewpoint conversion in S310, and a rule base 35a that has already been proven for many users already exists. It is suitable for the case of generating SVT. This is equivalent to selecting a rule adaptively.

  Furthermore, in the initial data setting process, a huge amount of work is required to individually set scoring vectors for all the rules (all items in the SVT 35b) described in the rule base 35a by the above-described method. Then, the following transfer process is executed.

  That is, if the attribute value for a certain situation attribute can be classified into several categories, the partial rule group consisting of rules whose attribute values for that situation attribute belong to the same category is assumed to have the same initial data. When a scoring vector is set for any rule belonging to a partial rule group, the same scoring vector is also transferred to other rules belonging to the partial rule group.

  Specifically, when the status attribute is “Destination”, “Pleasure Place” exists as one of the categories for the description of the destination, and “Park” and “Garden” are included in the “Pleasure Place” category. “Amusement park”, “resort”, “hot spring”, etc. are included. And there is a description about the status attribute “Destination” in the conditional clauses of Rule X and Rule Y, and the description is “park” in Rule X, “hot spring” in Rule Y, and the description about other status attributes Can be regarded as equivalent, the initial data of the scoring vectors for both rules X and Y are set to the same value. This is equivalent to analogy being executed in a sense.

  As described above, in the interface device 1 of the present embodiment, in the inference process, a plurality of rules are allowed to hit for a certain situation setting, and the scoring vectors of all the hit rules are cumulatively processed. Thus, a request vector necessary for obtaining the inference result is generated.

  Therefore, according to the interface apparatus 1 of the present embodiment, even if a plurality of rules are hit, an optimal inference result overlooking the hit rules can be obtained without falling into local optimization.

  In the present embodiment, since a plurality of rule groups are used for inference processing, and the description of the conditional clauses is made independent for each rule group, the conventional apparatus includes all the situation attributes in the conditional clauses. Compared to the above, it is possible not only to greatly reduce the number of rules to be set, but also to facilitate additional modification of the rules, and thus dynamic adaptation to individuals and situations.

  Further, in this embodiment, by describing the subordinate clause of a rule using a scoring vector, each rule can not only express the characteristic of the judgment result while leaving the ambiguity in the subordinate clause. And rules corresponding to special cases can be added easily. This addition may be added as a one-dimensional rule, or may be added as a conditional branch in a certain hierarchical rule. These correspond to adding a difference request vector under a certain condition to the request vector value of the basic case.

  Further, in the present embodiment, it is configured to select and use a rule group according to the situation from a plurality of rule groups, and only necessary rule groups are used in a limited manner. Can be executed adaptively and efficiently (and thus fast).

  In the present embodiment, the request attribute used for the scoring vector is set in the user's desire, and the attribute conversion process is performed after the scoring vector is accumulated, thereby converting the request attribute of the request vector V from the user's desire to the application. Therefore, when there are additions, changes, or deletions in the driveable application, it is possible to easily cope with changes in the transformation matrix P used for attribute conversion without changing the scoring vector. Can do.

  Further, in the present embodiment, the attribute value of the required attribute of the scoring vector is set based on the average user, and the viewpoint conversion process is performed after the attribute change process, so that the application vector is personalized by the user. It is designed to reflect the viewpoint.

  Therefore, the rule group (particularly the scoring vector) itself does not need to reflect the characteristics (viewpoint) of each user, and the same rule group can be applied to all users. The scale of the SVT 35b can be minimized.

  In the present embodiment, in the initial data setting process, scoring vectors are not individually set for all rules, but rules having attribute values belonging to the same category in a certain situation attribute are transferred by a transfer process. Since the same scoring vector is set, the processing amount in the initial data setting process can be greatly reduced.

  In the present embodiment, since the inference characteristics are corrected or strengthened based on the inference result and feedback information from the user, adaptation according to the individual and the situation can be dynamically performed.

  Further, in the present embodiment, the hierarchical rules are configured so that the description of the purpose and destination is given priority, so that the rules can be described efficiently. That is, when inferring a user request, the information of the destination / destination description is dominant. In particular, the destination description is preferable from the viewpoint of privacy protection because if the destination description is effectively used, the destination description can be prevented from being left directly on the terminal or the network.

  Incidentally, in the initial data setting process, it is necessary to prepare a template group of initial data in advance in order to realize the process (i) described above. In order to generate and edit this template group, it is desirable that the rule group can be visually represented on the display. At that time, the following expression method can be used.

  That is, two of the status attributes are associated with a two-dimensional address for individually specifying the tile images arranged two-dimensionally, and three of the request attributes are represented by the color of the tile image 3 Associate with primary colors.

  If the expression method is used in this way, the system designer or user can determine the characteristics of the rule group (such as the periodicity of the scoring vector and the spatial dependence) from the color distribution of the tile image, as shown in FIG. ) Visually. Specifically, it is possible to visually grasp a macro characteristic of a certain user, a similarity between different users, and a statistical characteristic common to many users.

  In addition, by imaging the characteristics of the rule group in this way, the rule group can be edited and processed (transfer of the rule within the same user / between other users) using a known image processing method performed on the display. Can be easily performed. That is, by introducing a scoring vector into the rule description, linear weighted sum processing between the rules can be easily performed. Therefore, such editing and processing, particularly, overlapping rule groups as shown in FIG. It is possible to combine them.

  Furthermore, by introducing a vector space or image expression into the expression of the rule group, it is possible to easily search and recognize the update history of the rule group and the rule (group) having a specific pattern.

  Note that two-dimensional addresses for specifying image tiles correspond to two conditions that are dominant in rule design among situation attributes, and the remaining situation attributes consist of a plurality of partial image tiles. The image may be configured to correspond to a relative address for specifying the partial image tile. In this case, more situation attributes can be reflected in a single image representing the characteristics of the rule group.

  In this embodiment, the rule base 35a and VST 35b are rule storage means, the situation storage table 35c is history storage means, S230 to S280 are scoring vector extraction means and request vector generation means, S120 is inference result generation means, and S290 is attribute conversion. Means S300 corresponds to viewpoint conversion means, S190 corresponds to initial data setting means and initial data transfer means, S170 corresponds to inference characteristic correction means and rule strengthening means, and S150 corresponds to response detection means.

Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment and can be implemented in various modes.
For example, in the above embodiment, by setting the conditional clause of the destination description or destination description at the higher level of the hierarchization rule, it is possible to describe the rule efficiently, but even when these are described as a one-dimensional rule Similar effects can be obtained. For example, if one destination description attribute such as {restaurant, post office, bank, company,...} Is defined, sufficient effects can be exhibited in information retrieval and navigation. The description of the time description by focusing on the season is one example in which multi-factor attributes are unified to enable efficient rule description.

  In the above embodiment, the request vector V obtained by accumulating the scoring vectors is adapted to individual users by performing attribute conversion and viewpoint conversion, but the contents of the situation data file and the user profile When selecting a rule group according to the above (S220), the SVT 35b may be adjusted (dynamic rewriting of the scoring vector) to be adapted to each user.

  In the above embodiment, the request attribute used for the scoring vector is set in the user's desire, and the attribute conversion process (S290) is performed after the scoring vector is accumulated, thereby converting the request attribute used for the request vector V into an application. However, the request attribute used for the scoring vector may be set in the application from the beginning, and the attribute conversion process (S290) may be omitted.

In the above embodiment, the viewpoint conversion is performed after the attribute conversion. However, the attribute conversion may be performed after the viewpoint conversion.
Moreover, in the said embodiment, although coordinate conversion is applied to the viewpoint conversion for every user, you may apply not only to this but to the dynamic situation adaptation in the same individual. That is, since the viewpoint changes according to various situations even for the same individual, the same model as in the case of viewpoint conversion can be used.

  In the above embodiment, the SVT 35b set based on the average user's viewpoint is used. However, when selecting the rule group (S220), the situation data and the contents of the user profile acquired in S210. Accordingly, the setting value (initial data) of the SVT 35b may be changed so as to be adapted to the current situation and the user. Since this is a process equivalent to the viewpoint conversion process in S300, in this case, the viewpoint conversion process in S300 may be omitted.

  In the above embodiment, the inference rules are described using the rule base 35a and the SVT 35b. However, the inference rules may be described in a programming language without using such a table. In this case, conditional clauses are described hierarchically by nesting of If statements.

  In the above embodiment, weighted addition is performed as the scoring vector accumulation process. However, other accumulation processes may be performed, and the calculation content may be linear or non-linear. In the above embodiment, every time a rule that hits is detected, the scoring vector corresponding to the rule is sequentially accumulated. However, the scoring vector of the hit rule is temporarily vectorized. The configuration may be such that the weighting factor is set and the accumulation process is executed after all the scoring vectors are extracted and stored in the list.

  In the above embodiment, the request vector V is converted into the application vector Va by performing the attribute conversion process. However, the request vector V and the application vector Va are clearly defined as in the expression (5). When the relationship is not defined, as shown in FIG. 12, it is necessary to represent the request vector V and the application vector Va in the same vector space. The space may be a request vector space whose coordinate axis corresponds to the user's desire, or a more general feature space. In this case, it is necessary to perform viewpoint conversion regarding the application in addition to viewpoint conversion regarding the user. Specifically, for example, in the coordinate system corresponding to the viewpoint of the user A, the similarity is calculated between the feature vectors F1 to FL of the application and the request vector VA of the viewpoint of the user A, and the degree of similarity is high. What is necessary is just to order an application in order.

It is a block diagram which shows the structure of the interface apparatus of embodiment. It is explanatory drawing which shows the structure of a rule base. It is explanatory drawing which shows the structure of a scoring vector table and a situation storage table. It is a functional block diagram of an interface apparatus. It is a flowchart which shows the content of the main process which a system control part performs. It is a flowchart which shows the content of inference processing. It is a schematic diagram which shows the structure of a rule group. It is explanatory drawing which shows the outline | summary of the accumulation process of a scoring vector. It is explanatory drawing which shows the specific example of an accumulation process. It is explanatory drawing which shows the content of the attribute conversion process. It is explanatory drawing which shows the content of a viewpoint conversion process. It is explanatory drawing which shows the modification of a viewpoint conversion process. It is explanatory drawing which shows the visual expression method of a rule group. It is explanatory drawing which shows the conventional rule structure.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Interface device, 3 ... Input device, 5 ... Display device, 7 ... Display, 8 ... Microphone, 11 ... Navigation device, 13 ... Air-conditioner device, 15 ... Audio device, 17 ... Communication device, 21 ... System control part, 23 ... Input I / F unit, 25 ... Display control unit, 27 ... Screen control unit, 28 ... Voice input unit, 29 ... Device control I / F, 31 ... Internet address database, 33 ... Search control unit, 35 ... Rule storage unit , 35a ... rule base, 35b ... scoring vector table (SVT), 35c ... situation storage table (SMT), 37 ... situation data storage unit, 37a ... situation data file, 37b ... user profile, 41-43 ... selection key, 44 ... operation key group, 51 to 53 ... assigned application display section, 54 ... provided application display section.

Claims (17)

An inference system that performs inference using rules expressed in an If-Then format consisting of conditional clauses and subordinate clauses,
An item set for classifying various situations is used as a situation attribute, and the conditional clause describes a linguistic or numerical condition for any one or a plurality of preset situation attributes. In addition, a scoring vector having a request attribute as an item representing a driveable application or an inferred user's desire and a requirement level for each of a plurality of preset request attributes as a subordinate clause is described. A rule storage means for storing a rule group comprising a plurality of rules,
Acquires status data indicating the status of the environment surrounding the user's situation and the user, based on the status data, satisfaction of conditional rules stored in the rule storage means, the non-fulfillment, constituting the condition clause Scoring that determines whether or not the description of the situation attribute matches the situation data, and extracts scoring vectors that constitute subordinate clauses of the rule for all rules that are judged to satisfy the conditional clause Vector extraction means;
By performing a cumulative process in which the scoring vector extracted by the scoring vector extraction means is weighted and added using a weighting factor set in advance in order to represent the importance of the scoring vector, it is possible to comprehensively calculate from the situation data. Request vector generation means for generating a request vector having as a component the degree of request for each of the request attributes in a situation that is grasped automatically,
Inference result generating means for generating an inference result based on the request vector generated by the request vector generating means;
An inference system characterized by comprising:   The inference according to claim 1, wherein the rule group stored in the rule storage means includes a one-dimensional rule in which a linguistic or numerical condition for a single situation attribute is described in the conditional clause. system.   The rule group stored in the rule storage means includes a hierarchical rule in which linguistic or numerical conditions for a plurality of situation attributes are described in the conditional clause. Inference system. The rule storage means stores a plurality of rule groups,
The inference system according to claim 1, wherein the scoring vector extraction unit selects a rule group to be used according to the situation data. The request vector generation means includes attribute conversion means for converting a request vector having the user's desire obtained by the accumulation process as the request attribute into a request vector having a driveable application as the request attribute. The inference system according to any one of claims 1 to 4 . The scoring vector is set to represent an average user characteristic, and the situation data includes a deviation of individual user characteristics relative to the average user characteristic and a coordinate transformation matrix. At least the viewpoint data expressed as a vector is included,
The request vector generation means, a request vector obtained by said cumulative, by conversion using the viewpoint data, billing, characterized in that it comprises the viewpoint converting means for generating a demand vector that reflects the characteristics of the user Item 6. The inference system according to any one of Items 1 to 5 . The initial data of the scoring vector aggregates evaluation attribute value based on the subjective judgment of a plurality of users, any of claims 1 to 6, characterized in that obtained from the results averaged statistically Inference system described in Crab. The inference system according to claim 1 , wherein the initial data of the scoring vector is a numerical value of a subjective empirical rule of a specific user. The inference system according to claim 1 , further comprising initial data setting means for setting initial data of the scoring vector. Said initial data setting means, an input data from outside, or the template selected by the external operation from a set of predefined routine template group, in claim 9, characterized in that to set as the initial data The inference system described. The initial data setting means includes
A part of the situation attribute included in the rule is set as a partial situation attribute, and the rule group is classified into a plurality of partial rule groups for each rule in which a condition related to the same partial situation attribute is described in the conditional clause. When initial data of a scoring vector is set for, an initial data transfer means for transferring the same initial data to a scoring vector of another rule belonging to the same partial rule group as the rule is provided. The inference system according to claim 9 or 10 . The request vector used when the inference result generation means generates the inference result and the response data indicating the response of the user who has received the inference result are used as feedback information for the inference result.
It is set in advance between an operation vector generated from the response data and representing a user's actual request in the same format as the request vector, and a request vector associated with the response data by the feedback information. When there is a difference greater than the threshold value, or when it is statistically recognized that there is a difference greater than the threshold value between the two vectors from the history of the feedback information within a preset fixed period The target rule is such that the difference between the operation vector and the request vector is reduced based on the target rule that is the rule used to generate the request vector that satisfies the correction condition. inference characteristics modifier hand by changing the value of scoring vectors constituting a subordinate clause modifies the inference characteristics of the system Inference system according to any one of claims 1 to 11, characterized in that it comprises a. The inference system according to claim 12 , wherein the feedback information relates to the same user. The inference system according to claim 12, wherein the feedback information is related to another user. The inference system according to claim 12 , further comprising history storage means for storing a history of the feedback information. Response detection means for detecting a response from the user who has received the inference result generated by the inference result generation means;
When the response detected by the response detection means affirms the inference result, the rule related to the request vector used for obtaining the inference result is set as the strengthening target rule, and the strengthening target rule is strengthened. Rules strengthening means,
Equipped with a,
The rule strengthening means changes the weighting factor used when the request vector is generated from the scoring vector constituting the subordinate clause of the rule to be strengthened in the demand vector generating means, or the rule to be strengthened The inference system according to claim 1, wherein the rule to be strengthened is strengthened by increasing a value of a scoring vector that constitutes a subordinate clause . The rule strengthening means is a rule in which a condition related to the same partial status attribute is described in the conditional clause, with a part of the status attribute included in the strengthened rule as a partial status attribute. The inference system according to claim 16 , which is also an enhancement target.