JP2012212276A - Learning unit, learning method and learning program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a learning unit capable of curbing an excessive increase or decrease in learning sensitivity, in a Bayesian network that updates a conditional probability list by learning based on observation.SOLUTION: A learning unit 10 of an inference device 100 comprises: an observation frequency increase-decrease management section 111 for adding an observation frequency addition value to an observation frequency of a cell relevant to observation data in a conditional observation frequency list 23, based on the observation data including each observation value of parent and child nodes; a total observation frequency calculation section 112 for calculating a total of observation frequencies in the conditional observation frequency list 23; an observation frequency list normalization section 113 for, when the total of observation frequencies exceeds a predetermined upper limit, normalizing the conditional observation frequency list 23 so that the total of observation frequencies becomes equivalent to a predetermined reference value lower than the upper limit; and a conditional probability list update section 12 for updating a conditional probability list 22 corresponding to the conditional observation frequency list 23.

Description

  The present invention relates to a learning apparatus, a learning method, and a learning program for updating a conditional probability table of a Bayesian network by learning based on observation.

  Conventionally, a Bayesian network is known as an information processing model that performs probability calculation based on the dependency of a plurality of factors. A Bayesian network is a network-like probability model defined by a random variable, a conditional dependency between random variables, and a conditional probability.

  A random variable is represented by a node, and a conditional dependency between random variables is represented by a directed link extending between the nodes. A node that comes before the link is called a child node, and a node that is at the link source is called a parent node. The conditional probability is a probability that a child node takes a certain value when the parent node takes a certain value. In the case of a discrete variable, P (child node = y | parent node = x1, x2,...) = P is expressed in the form of a table (conditional probability table) in which the probability values in all the states taken by the child node and the parent node are listed (Non-Patent Document 1).

  FIG. 3 is a diagram showing a graph structure of a Bayesian network in which a directed link is extended from a parent node “scene” and “total program time” to a child node “genre”, respectively. It is a figure which shows the conditional probability table | surface of a Bayesian network. As shown in FIG. 4, in the conditional probability table, probability values of “genre” are listed for each “scene” and “total program time”.

  The conditional probability table shown in FIG. 4 is created based on the conditional observation frequency table as shown in FIG. In other words, the conditional probability is a state of a parent node (in the example of FIG. 3, “scene” and “total program time”, which corresponds to the column in FIG. 4), and each state of the child node (example in FIG. 3). Is a “genre” and corresponds to the row in FIG. 4), each probability is the corresponding cell in the conditional observation count table when the observation values of the parent node and the child node are obtained. And the number of observations of each state of the child node in the state of the parent node is divided by the total number of observations of the state of the parent node, respectively, so that each state of the child node in the state of the parent node Can be obtained.

  Specifically, as shown in FIG. 5, for example, in the state of the parent node “scene” = “leisure” and “total program time” = “2-2.5 hours”, the “genre” of the child node The number of observations is “J-POP” 5 times, “Rock” 3 times, “Jazz” 11 times, “Classic” 1 time, “Folk” 3 times, “Dance” 1 time, The total number of observations is 24.

  Therefore, in the conditional probability table of FIG. 4, regarding the state of the parent node “scene” = “leisure” and “total program time” = “2-2.5 hours”, the child node “genre” is conditionalized. Probability is “J-POP” 5 times / 24 times = 20.8%, “Rock” 3 times / 24 times = 12.5%, “Jazz” 11 times / 24 times = 45.8%, “Classic” is 1 time / 24 times = 4.2%, “Folk” is 3 times / 24 times = 12.5%, and “Dance” is 1 time / 24 times = 4.2%.

  When the conditional probability table is updated by learning based on observation, first, the conditional observation number table is updated by increasing the number of observations of the corresponding cell in the conditional observation number table shown in FIG. Then, the conditional probability table is updated by recreating the conditional probability table based on the updated conditional observation frequency table.

  For example, when observation data is obtained that “scene” = “leisure” and “total program time” = “2-2.5 hours” and “genre” is “lock”, conditional observation In the frequency table, “Scene” = “Leisure” and “Program total time” = “2 to 2.5 hours”, and “Genre” = “Lock” is added to the observation frequency addition value. The conditional probability table is updated by recalculating the conditional probability table based on the conditional observation number table after the addition of the observation number addition value.

  In this way, by updating the conditional probability table by learning based on observation, the reliability of the conditional probability table is improved, and the accuracy of inference using the conditional probability table can be improved.

Yoichi Motomura “Bayesian Net Software”, Journal of the Japanese Society for Artificial Intelligence, Vol. 17 No. 5a (2002)

  However, as learning based on observation is repeated, the total number of observations in the conditional observation table increases. When the total number of observations becomes too large, the learning effect (learning sensitivity) by one observation becomes too small, and the conditional probability table is hardly updated by learning by one observation.

  In addition, learning may be performed to increase or decrease the number of observations in the conditional observation number table based on the observation. In such a learning method, by repeating learning based on observation, the number of observations in the conditional observation number table may decrease and the total number of observations may decrease. If the total number of observations becomes too small, the learning effect (learning sensitivity) by one observation becomes too high, and the conditional probability table greatly fluctuates by learning by one observation.

  In addition, if the number of observations in the observation number table decreases and all the observation times become zero, the conditional probability cannot be calculated, and the Bayesian network itself cannot be calculated.

  The present invention has been made in view of the above problems, and in a Bayesian network that updates a conditional probability table by learning based on observation, a learning device that can suppress an excessive increase or excessive decrease in learning sensitivity. The purpose is to provide.

  The learning device of the present invention is a learning device that updates a conditional probability table of a Bayesian network by learning based on observation, and based on observation data including observation values of a parent node and observation values of a child node, the parent node An observation frequency increase / decrease management unit that adds the observation frequency addition value to the observation frequency of the corresponding combination in the conditional observation frequency table that defines the observation frequency for each combination of each state of the child node and each state of the child node, and the conditional An observation number sum calculation unit for calculating the total number of observations in the observation number table, and when the total number of observations exceeds a predetermined upper limit value, a predetermined reference value in which the total number of observation times is lower than the upper limit value; An observation frequency table normalization unit that normalizes the conditional observation frequency table, and a conditional confirmation corresponding to the conditional observation frequency table based on the conditional observation frequency table. It has a configuration that includes a conditional probability table updating unit that updates a table.

  According to this configuration, the conditional observation frequency table is normalized so that the total number of observations becomes the reference value when the total number of observations exceeds a predetermined upper limit value. Can be suppressed.

  In the learning device of the present invention, the observation frequency increase / decrease management unit may increase or decrease the observation frequency addition value based on the observation data.

  According to this configuration, even when the increase in the number of observations is accelerated based on the observation data, an excessive decrease in learning sensitivity can be suppressed.

  A learning device according to another aspect of the present invention is a learning device that updates a conditional probability table of a Bayesian network by learning based on observation, and is based on observation data including observation values of a parent node and observation values of a child node. , Adding the observation count addition value to the number of observations in the conditional observation count table that defines the number of observations for each combination of each state of the parent node and each state of the child node, or the number of observations The observation frequency increase / decrease management unit for subtracting the subtraction value, the observation frequency total calculation unit for calculating the total number of observations in the conditional observation frequency table, and when the total number of observations exceeds a predetermined upper limit value, or When the total number of observations falls below a predetermined lower limit value, the condition is set so that the total number of observations becomes a predetermined reference value smaller than the upper limit value and larger than the lower limit value. An observation frequency table normalization unit that normalizes the observation frequency table, and a conditional probability table update unit that updates the conditional probability table corresponding to the conditional observation frequency table based on the conditional observation frequency table. It has a configuration.

  According to this configuration, when the total number of observations exceeds a predetermined upper limit value, or when the total number of observations falls below a predetermined lower limit value, the total number of observations is set to the reference value. Since the table is normalized, an excessive increase and excessive decrease in learning sensitivity can be suppressed.

  In the learning device of the present invention, the observation frequency increase / decrease management unit may increase or decrease the observation frequency addition value or the observation frequency subtraction value based on the observation data.

  According to this configuration, even when the increase and decrease in the number of observations are accelerated based on the observation data, it is possible to suppress the excessive increase and excessive decrease in the learning sensitivity.

  In the above learning apparatus, the Bayesian network may be a probability model related to reproduction of information content, and includes a node related to the attribute of the information content and a node related to the attribute of the reproduction, and a node related to the attribute of the information content The number of observations may increase or decrease based on the observation data.

  According to this configuration, since the number of observations increases or decreases for the nodes related to the information content attribute based on the observation data, the learning sensitivity is excessively increased due to excessive increase and decrease of the nodes related to the information content attribute. Increase and excessive reduction can be suppressed. Note that the information content may be, for example, music content, and the node related to the attribute of the information content may include the genre of the music, the age, the artist, and the presence of each information content. In addition, the node relating to the reproduction of the information content may include a reproduction day of the week, a time, a scene, and a total program time.

  In the learning device of the present invention, the information content can be automatically reproduced in a predetermined order and can be directly reproduced. The observation frequency increase / decrease management unit reproduces the information content. When the information content is reproduced, the observation number addition value is added to the number of observations of the corresponding state of the node related to the information content attribute. The state observation count addition value may be larger than the observation count addition value of the corresponding state of the node related to the attribute of the information content when automatically reproduced.

  According to this configuration, the increase in the number of observations can be accelerated when the information content is directly reproduced.

  In the learning device of the present invention, the information content can be automatically reproduced according to a predetermined order and can be repeatedly reproduced. The observation frequency increase / decrease management unit reproduces the information content. When the information content is repeatedly played, the corresponding state of the node related to the information content attribute is added to the number of observations of the corresponding state of the node related to the information content attribute. May be larger than the observation count addition value of the corresponding state of the node related to the attribute of the information content when automatically reproduced.

  According to this configuration, the increase in the number of observations can be accelerated when the information content is repeatedly reproduced.

  The information content can be automatically reproduced according to a predetermined order, and can be skipped during automatic reproduction according to the order. When the reproduction of the content is skipped, the observation number subtraction value may be subtracted from the observation number of the corresponding state of the node related to the information content attribute.

  According to this configuration, when the reproduction of information content is skipped, the number of observations can be reduced.

  Yet another aspect of the present invention is a learning method for updating a conditional probability table of a Bayesian network by observation-based learning, which includes observation data including observation values of a parent node and observation values of a child node. Based on the observation number increase / decrease management for adding the observation number addition value to the number of observations of the combination in the conditional observation number table that defines the number of observations for each combination of each state of the parent node and each state of the child node Step, a total number of observations calculation step for calculating the total number of observations in the conditional observation number table, and when the total number of observations exceeds a predetermined upper limit, the total number of observations is greater than the upper limit. Based on the observation frequency table normalizing step for normalizing the conditional observation frequency table so as to be a low predetermined reference value, and the conditional observation frequency table, And a probability table updating step conditional updating conditional probability table corresponding to the conditional number of observations table.

  Even with this configuration, when the total number of observations exceeds a predetermined upper limit, the conditional observation table is normalized so that the total number of observations becomes the reference value. Can be suppressed.

  Still another aspect of the present invention is a learning method for updating a conditional probability table of a Bayesian network by observation-based learning, wherein the observation method includes observation values of a parent node and observation values of a child node. Add the observation count addition value to the observation count of the corresponding combination in the conditional observation count table that defines the number of observations for each combination of each state of the parent node and each of the child nodes, or subtract the observation count value The observation frequency increase / decrease management step for subtracting, the observation frequency summation step unit for calculating the total number of observations in the conditional observation frequency table, and when the total of the observation frequency exceeds a predetermined upper limit, or the observation When the total number of times falls below a predetermined lower limit value, the total number of observations becomes a predetermined reference value smaller than the upper limit value and larger than the lower limit value, An observation frequency table normalizing step for normalizing the conditional observation frequency table, and a conditional probability table updating step for updating the conditional probability table corresponding to the conditional observation frequency table based on the conditional observation frequency table Including.

  Even with this configuration, when the total number of observations exceeds a predetermined upper limit value or when the total number of observations falls below a predetermined lower limit value, the observation number table is set so that the total number of observations becomes the reference value. Is normalized, so that an excessive increase and excessive decrease in learning sensitivity can be suppressed.

  Yet another embodiment of the present invention is a computer program for causing a computer to execute the learning method described above.

  According to the present invention, in a Bayesian network in which a conditional probability table is updated by learning based on observation, an excessive increase or excessive decrease in learning sensitivity can be suppressed.

The block diagram which shows the structure of the learning apparatus in embodiment of this invention The figure which shows the graph structure of the conditional dependence relationship in embodiment of this invention The figure which shows the graph structure of a part of conditional dependency in embodiment of this invention The figure which shows the example of the conditional probability table | surface in embodiment of this invention The figure which shows the example of the conditional observation frequency table in embodiment of this invention Flow chart of conditional probability update processing according to the embodiment of the present invention

  Hereinafter, a learning apparatus according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing a configuration of an inference device including a learning device according to an embodiment of the present invention. The inference apparatus 100 is an apparatus that performs inference using a probability model related to reproduction of music content that is information content.

  The inference device 100 is connected to an information content reproduction device (music content reproduction device) not shown, and receives observation data for learning a conditional probability table from the information content reproduction device. In addition, the inference device 100 inputs an observation value of a node (hereinafter referred to as an “observation node”) from which an observation value is obtained from the information content reproduction device when performing inference, and a node to be inferred (hereinafter referred to as “observation node”) The inference result of “inference node” is output to the information content reproduction apparatus.

  As illustrated in FIG. 1, the inference apparatus 100 according to the present embodiment includes a learning unit 10, a model storage unit 20, an inference unit 30, an input unit 40, and an output unit 50. The learning unit 10 corresponds to the learning device of the present invention. The learning unit 10 may be realized by a computer executing a learning program.

  The model storage unit 20 stores a Bayesian network probability model. The learning unit 10 updates the probability model stored in the model storage unit 20 by learning. The inference unit 30 creates an a posteriori probability table of the inference node based on the observation value of the observation node. The inference unit 30 may output the posterior probability table as an inference result, or may output the state of the maximum posterior probability in the posterior probability table as the inference result. The input unit 40 inputs the observation value of the parent node and its child node from the information content reproduction device during learning, and inputs the observation value of the observation node from the information content reproduction device during inference. The output unit 50 outputs the inference result obtained by the inference unit 30.

  Specifically, the model storage unit 20 includes a conditional dependency relationship 21 between random variables (that is, a Bayesian network graph structure), a conditional probability table 22, and a conditional probability table 22 for creating the conditional probability table 22. The observation frequency table 23 is stored.

  FIG. 2 is a diagram illustrating a graph structure of the conditional dependency relationship 21. As described above, this probability model is a probability model related to reproduction of music content as information content. As shown in FIG. 2, this probability model includes nodes of “genre”, “age”, “artist”, and “presence / absence of each information content” as nodes related to the attributes of the music content, As nodes related to the reproduction of the music content, the nodes of “Day of Week”, “Time”, “Scene”, and “Total Program Time” of reproduction are included.

  When a certain event is observed, the input unit 40 inputs the value of a random variable as observation data of the event in order to update the conditional observation count table 23. The input unit 40 inputs observation data from the information content reproduction device. This information content reproduction apparatus is mounted on a vehicle.

  When the music content is played back, the information content playback device outputs the values of “genre”, “age”, “artist”, and “presence / absence of each information content” as the observed values related to the attributes of the music content, and plays them. As the observed values, the values of “day of the week”, “time”, “scene”, and “total program time” of the reproduction are output. The input unit 40 inputs these values as observation data.

  The information content reproduction apparatus can automatically reproduce a plurality of music contents according to a predetermined reproduction order. In addition, the information content reproduction apparatus can also perform direct reproduction in which music content to be reproduced is directly designated and reproduced in accordance with a user operation. Furthermore, the information content reproduction apparatus can perform repeat reproduction in which a plurality of music contents are not reproduced in order according to the user's operation in accordance with the reproduction order, but one music content is repeatedly reproduced.

  In addition, the information content playback apparatus can also skip the one music content and play the next music content when the plurality of music contents are played in order according to the playback order according to the user's operation. In addition to the above-mentioned observation data at the time of reproduction, the information content reproduction device outputs direct reproduction information, repeat reproduction information, and reproduction skip information as observation data, and the input unit 40 inputs these observation data. To do.

  Hereinafter, learning of the conditional probability table will be described. In the following description, among the graph structure of the conditional dependency relationship 21 shown in FIG. 2, graphs relating to “scene”, “total program time”, and “genre” The structure portion will be described as an example. FIG. 3 is a diagram showing a conditional dependency relationship related to the graph structure portion. In this graph structure portion, “scene” and “total program time” are parent nodes, and “genre” is a child node.

  FIG. 4 is a diagram showing an example of the conditional probability table 22 of this graph structure part. In this conditional probability table, each column represents each state of “scene” as a parent node and “total program time”, and each row represents each state of “genre” as a child node. FIG. 5 is a diagram illustrating an example of the conditional observation count table 23. The conditional observation frequency table in FIG. 5 corresponds to the conditional probability table in FIG.

  The learning unit 10 includes a conditional observation frequency elevation unit 11 including a conditional observation frequency update unit 11, an observation frequency total calculation unit 112, an observation frequency table normalization unit 113, and a conditional probability update unit 12. I have.

  The observation frequency increase / decrease management unit 111 acquires from the input unit 40 observation data including observation values of the parent node “scene” and “program total time” and child node “genre” observation values. Based on the observation node, the observation frequency increase / decrease management unit 11 sets each state of the “scene” as a parent node (leisure, commuting, shopping, etc.) and each state of the “total program time” as a parent node (from 0.5 to 0.5). 1 hour, 1 to 1.5 hours, 1.5 to 2 hours, ...) and each combination of child node "genre" states (J-POP, rock, jazz, classic, folk, dance) In the conditional observation count table in which the number of observations is defined, the observation count addition value is added to the observation count of the cell corresponding to the acquired observation node.

  The observation frequency increase / decrease management unit 111 adds the reference observation frequency addition value for the music content that is automatically reproduced according to the reproduction order. The reference observation frequency addition value is, for example, 1. The observation frequency increase / decrease management unit 111 performs positive learning that adds the additional observation frequency addition value for the music content that is not directly reproduced according to the reproduction order but is directly designated and reproduced by the user's operation. . In addition, the observation frequency increase / decrease management unit 111 performs positive learning for adding the additional observation frequency addition value in the same manner for the music content that is repeatedly played by the user's operation. The additional observation count addition value is 2, for example.

  The observation frequency increase / decrease management unit 111 not only increases the observation frequency based on the observation data, but also performs negative learning to decrease the observation frequency in the following case. That is, in the information content reproduction apparatus, when the music content is skipped, the observation frequency increase / decrease management unit 111 acquires the information as observation data, and for the skipped music content, the observation frequency of the corresponding cell. Subtract the observation count subtraction value from. The observation frequency subtraction value is, for example, 1.

  The observation count total calculation unit 112 calculates the sum of the observation counts in the conditional observation count table 23. The observation number table normalization unit 113 determines whether the total number of observations calculated by the observation number total calculation unit 112 exceeds a predetermined upper limit value (for example, 1200).

  When the total number of observations exceeds the upper limit, the observation number table normalization unit 113 sets the number of conditional observations so that the total number of observations becomes a predetermined reference value (for example, 1000) lower than the upper limit. Normalize the table. Specifically, the number of observations of all cells is multiplied by (reference value / upper limit value) (in the above example, 1000/1200 = 5/6) so that the total number of observations is about the reference value. . That is, the total number of observations calculation unit 112 sets all the observation numbers so that the total number of observations is returned to the reference value when the total number of observations exceeds the reference value by a predetermined ratio (20% in the above example). Multiply the number of cell observations by (reference value / upper limit) = (1 / (1 + predetermined ratio)) to normalize the entire conditional observation table.

  In addition, the observation number table normalization unit 113 determines whether or not the total number of observations calculated by the observation number total calculation unit 112 is below a predetermined lower limit value (for example, 800). When the total number of observations falls below the lower limit, the observation number table normalization unit 113 normalizes the conditional observation number table so that the total number of observations becomes a reference value higher than the lower limit (for example, 1000). Turn into.

  Specifically, the number of observations of all cells is multiplied by (reference value / lower limit value) (10000/800 = 5/4 in the above example) so that the total number of observations is about the reference value. . That is, the total number of observations calculation unit 112 sets all the observation numbers so that the total number of observations is returned to the reference value when the total number of observations falls below the reference value by a predetermined ratio (20% in the above example). Multiply the number of cell observations by (reference value / lower limit value) = (1 / (1−predetermined ratio)) to normalize the entire conditional observation frequency table.

  As described above, the conditional observation count table updating unit 11 updates the conditional observation count table 23 and normalizes the conditional observation count table 23 so that the sum of the observation counts becomes a reference value as necessary. To do. The conditional probability table updating unit 12 recreates the conditional probability table based on the conditional observation frequency table obtained in this way.

  In the above, the observation data input by the input unit 40 when performing learning based on observation in the learning unit 10 has been described. However, the input unit 40 is further configured so that the inference unit 30 determines the inference node based on the value of the observation node. In order to make an inference, the observation value of the observation node is input.

  The inference unit 30 performs inference using a Bayesian network probability model stored in the model storage unit 20 based on the value of the observation node input to the input unit 40. 2, the inference unit 30 uses the “scene” and “total program time” nodes as observation nodes, and also includes “genre”, “age”, “artist”, and “information contents”. Inference is performed using “existence” as an inference node.

  The inference unit 30 generates a posterior probability table of the observation node based on the observation value of the observation node, and obtains a posterior probability table of the estimation node based on the posterior probability table of the observation node. The inference unit 30 outputs the posterior probability table of the inference node to the information content reproduction device via the output unit 50. The information content reproduction apparatus presents recommended music content in consideration of the user's tendency and preference based on the a posteriori probability table of the inference node. The inference unit 30 may output the music content having the maximum posterior probability in the posterior probability table of the inference node to the information content reproduction device via the output unit 50.

  FIG. 6 is a flowchart of conditional probability update processing by the learning unit 10. The learning unit 10 first acquires observation data from the input unit 40 (step S61). Whether the observation frequency increase / decrease management unit 111 of the conditional observation frequency table update unit 11 selects a cell for updating the observation frequency in the conditional observation frequency table 23 based on the observation data, and increases the observation frequency of the cell. Decide whether to decrease, and if you want to increase it, determine whether to add the standard observation count addition value or the additional observation count addition value, and count the number of observations relative to the observation count of the selected cell. The addition value is added or the observation number subtraction value is subtracted (step S62).

  Next, the observation number total calculation unit 112 calculates the total number of observations for the conditional observation number table 23 (step S63). Then, the observation number table normalization unit 113 has the total number of observations calculated by the observation number sum calculation unit 112 exceeding a predetermined upper limit value (20% increase of the reference value), or a predetermined lower limit value. It is determined whether it is below (20% reduction of the reference value) (step S64).

  When the total number of observations exceeds a predetermined upper limit value or is lower than a predetermined lower limit value (YES in step S64), observation number table normalization unit 113 stores conditional observation number table 23. Normalization is performed (step S65). If the total number of observations is less than or equal to the upper limit value and greater than or equal to the lower limit value (NO in step S64), and after normalizing the conditional observation frequency table 23 by the observation frequency table normalization unit 113, the conditional The probability table update unit 12 updates the conditional probability table based on the conditional observation count table 23 (step S66).

  The conditional probability table may be updated any time after the number of observations in the observation number table is increased or decreased in step S62. This is because normalization of the conditional observation count table 23 is performed by multiplying the whole by a constant coefficient as described above, so that the conditional probability table 22 calculated based on the conditional observation count table 23 is This is because the conditional observation count table 23 does not change before and after normalization.

  In the above embodiment, the upper limit value is increased by 20% of the reference value and the lower limit value is decreased by 20% of the reference value. However, it goes without saying that the upper limit value and the lower limit value are not limited to these. . In the above embodiment, the reference value is 1000, but the present invention is not limited to this. The reference value is set based on the number of cells, the number of observation count addition values, and the number of observation count subtraction values. Preferably, the reference value is determined so that the conditional probability of the conditional probability table varies by about 0.1 to 10% by adding the observation number addition value for one time.

  In the above embodiment, when the total number of observations in the conditional observation count table 23 exceeds the upper limit value, all the observation counts in the conditional observation count table 23 are multiplied by (reference value / upper limit value). Or, when the value is below the lower limit value, the conditional observation count table 23 is normalized by multiplying all the observation counts in the conditional observation count table 23 by (reference value / lower limit value). When the total number of observations in Table 23 exceeds the upper limit value, multiply all the observation times in the conditional observation number table 23 by (reference value / total number of observations) or below the lower limit value. The conditional observation count table 23 may be normalized by multiplying all the observation counts in the additional observation count table 23 by (reference value / total number of observations).

  According to the latter method, the sum of the number of observations becomes the reference value just by normalization. However, since it is not necessary to accurately return the total number of observations to the reference value during normalization, even the former method does not substantially affect the effects of the present invention.

  In the above embodiment, the observation frequency increase / decrease management unit 111 increases or decreases the observation frequency based on the observation data. However, when the observation data is obtained, the corresponding cell in the conditional observation frequency table 23 It is also possible to simply add the observation count addition value to. In addition, the observation number increase / decrease management unit 111 adds the additional observation number addition value to the corresponding cell in the conditional observation number table 23 in order to accelerate the addition of the observation number when direct reproduction or repeat reproduction is performed. However, it is not necessary to add such observation times. On the other hand, the observation frequency increase / decrease management unit 111 also accelerates the subtraction of the observation frequency by subtracting the additional observation frequency subtraction value when a more negative event is observed in the subtraction of the observation frequency. Good.

  The present invention has an effect of suppressing an excessive increase or decrease in learning sensitivity in a Bayesian network that updates a conditional probability table by learning based on observation, and observes a conditional probability table of a Bayesian network. It is useful as a learning device that is updated by learning based on the above.

DESCRIPTION OF SYMBOLS 10 Learning part 11 Conditional observation frequency table update part 111 Observation frequency increase / decrease management part 112 Observation frequency sum total calculation part 113 Observation frequency table normalization part 12 Conditional probability update part 20 Model storage part 21 Conditional dependence 22 Conditional probability table 23 Conditional Observation Count Table 30 Inference Unit 40 Input Unit 50 Output Unit 100 Inference Device

Claims (11)

A learning device that updates a conditional probability table of a Bayesian network by observation-based learning,
Based on the observation data including the observation value of the parent node and the observation value of the child node, the corresponding observation number in the conditional observation number table defining the number of observations for each combination of each state of the parent node and each state of the child node An observation frequency increase / decrease management unit for adding the observation frequency addition value to the number of observations of the combination;
A total number of observations calculation unit for calculating the total number of observations in the conditional number of observations table;
An observation frequency table normalization that normalizes the conditional observation frequency table so that the total number of observations becomes a predetermined reference value lower than the upper limit value when the total number of observations exceeds a predetermined upper limit value. And
A conditional probability table updating unit that updates a conditional probability table corresponding to the conditional observation frequency table based on the conditional observation frequency table;
A learning apparatus comprising:   The learning apparatus according to claim 1, wherein the observation frequency increase / decrease management unit increases or decreases the observation frequency addition value based on the observation data. A learning device that updates a conditional probability table of a Bayesian network by observation-based learning,
Corresponding combinations in the conditional observation count table that specifies the number of observations for each combination of each state of the parent node and each state of the child node based on observation data including the observation value of the parent node and the observation value of the child node The observation frequency increase / decrease management unit for adding the observation frequency addition value or subtracting the observation frequency subtraction value to the observation frequency of
A total number of observations calculation unit for calculating the total number of observations in the conditional number of observations table;
When the total number of observations exceeds a predetermined upper limit value or when the total number of observations falls below a predetermined lower limit value, the total number of observations is smaller than the upper limit value and larger than the lower limit value. An observation frequency table normalization unit that normalizes the conditional observation frequency table so as to be a reference value;
A conditional probability table updating unit that updates a conditional probability table corresponding to the conditional observation frequency table based on the conditional observation frequency table;
A learning apparatus comprising:   The learning device according to claim 3, wherein the observation frequency increase / decrease management unit increases or decreases the observation frequency addition value or the observation frequency subtraction value based on the observation data.   The Bayesian network is a probabilistic model related to reproduction of information content, and includes a node related to the attribute of the information content and a node related to the attribute of the reproduction, and the node related to the attribute of the information content is based on the observation data The learning apparatus according to claim 1, wherein the number of observations increases or decreases. The information content can be automatically reproduced according to a predetermined order, and can also be directly reproduced.
When the information content is reproduced, the observation frequency increase / decrease management unit adds the observation frequency addition value to the observation frequency of the corresponding state of the node related to the attribute of the information content, and when the information content is directly reproduced. In addition, the observation number addition value of the corresponding state of the node related to the attribute of the information content is made larger than the observation number addition value of the corresponding state of the node related to the attribute of the information content when the information content is automatically played back. The learning device according to claim 5. The information content can be automatically reproduced according to a predetermined order, and can also be reproduced repeatedly.
When the information content is reproduced, the observation frequency increase / decrease management unit adds the observation frequency addition value to the observation frequency of the corresponding state of the node related to the attribute of the information content, and when the information content is reproduced repeatedly In addition, the observation number addition value of the corresponding state of the node related to the attribute of the information content is made larger than the observation number addition value of the corresponding state of the node related to the attribute of the information content when the information content is automatically played back. The learning device according to claim 5 or 6. The information content can be automatically reproduced according to a predetermined order, and can be skipped during automatic reproduction according to the order,
The observation frequency increase / decrease management unit subtracts the observation frequency subtraction value from the observation frequency of a corresponding state of a node related to the attribute of the information content when reproduction of the information content is skipped. The learning apparatus as described in any one of thru | or 7. A learning method for updating a conditional probability table of a Bayesian network by observation-based learning,
Based on the observation data including the observation value of the parent node and the observation value of the child node, the corresponding observation number in the conditional observation number table defining the number of observations for each combination of each state of the parent node and each state of the child node Observation number increase / decrease management step of adding the observation number addition value to the number of observations of the combination;
A total number of observations calculation step for calculating the total number of observations in the conditional observation number table;
An observation frequency table normalization that normalizes the conditional observation frequency table so that the total number of observations becomes a predetermined reference value lower than the upper limit value when the total number of observations exceeds a predetermined upper limit value. Step,
A conditional probability table update step for updating a conditional probability table corresponding to the conditional observation frequency table based on the conditional observation frequency table;
The learning method characterized by including. A learning method for updating a conditional probability table of a Bayesian network by observation-based learning,
Corresponding combinations in the conditional observation count table that specifies the number of observations for each combination of each state of the parent node and each state of the child node based on observation data including the observation value of the parent node and the observation value of the child node The observation frequency increase / decrease management step of adding the observation frequency addition value or subtracting the observation frequency subtraction value to the observation frequency of
A total number of observations calculation step for calculating the total number of observations in the conditional observation number table;
When the total number of observations exceeds a predetermined upper limit value or when the total number of observations falls below a predetermined lower limit value, the total number of observations is smaller than the upper limit value and larger than the lower limit value. An observation frequency table normalizing step for normalizing the conditional observation frequency table to be a reference value;
A conditional probability table update step for updating a conditional probability table corresponding to the conditional observation frequency table based on the conditional observation frequency table;
The learning method characterized by including.   A learning program for causing a computer to execute the learning method according to claim 9 or 10.

Images