JP3151857B2 - Inference device with learning function - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an inference apparatus suitable for a system targeting a nonlinear model such as a stock price prediction system, and more particularly to an inference apparatus having a learning function.

[0002]

2. Description of the Related Art A conventional inference apparatus with a learning function is disclosed, for example, in Japanese Patent Application Laid-Open No. 2-178801.
In the inference apparatus disclosed in the publication, fuzzy is used for inference. In this apparatus, the learning for inference is started every predetermined period, for example, the number of times of data input is counted, and every time the number of times exceeds a predetermined number of times. As a result, for example, when the present invention is applied to controls such as the volume of a television and the temperature of a cooler, restrictions are gradually made in accordance with the preference of the operator.

[0003]

However, the above-described inference apparatus has the following problems.

(1) It is impossible to designate a start time and an end time of learning corresponding to a change in state of an inference target.

(2) Even when the state of the inference target changes abruptly and the inference result of the knowledge group up to that point loses validity, learning cannot be started until a predetermined number of inputs is reached.

That is, when the inference target is a non-linear model such as a stock price prediction system, there is a disadvantage that the learning function of the conventional inference apparatus as described above hardly provides a learning effect.

An object of the present invention is to provide an inference apparatus with a learning function capable of solving the above-mentioned drawbacks by performing learning in response to a state change of an inference target.

[0008] The present invention receives the input value, and knowledge group inference means for performing inference receives inference result and legal value of the knowledge group inference means, to evaluate the need for learning for knowledge group update, learning starting the learning function activation determining means for determining the timing of the completion of the learning in the learning function activation determining means
A knowledge group inference unit updating means for learning a knowledge group from an input value and a valid value when it is determined necessary and updating the knowledge group in the knowledge group inference means .
In the inference apparatus, the learning function activation determining means includes:
Inference conclusion of each knowledge group of the knowledge group inference means for each logic cycle
The result of comparing the result with the legal value is used as the evaluation value of each knowledge group.
Infer each of the past m times (m> 1) memorized and stored
Infer the need for learning based on the evaluation value for the cycle
Judgment is made for each vehicle and the timing of learning start and end is determined
It is characterized by doing.

The knowledge group inference unit updating means includes: a knowledge group learning means for creating a learned knowledge group as a new knowledge group;
Means for adding a new knowledge group to the knowledge group inference means.

[0010]

[0011]

A plurality of knowledge groups are usually provided in parallel in the knowledge group inference means, and input values are input to these knowledge groups in parallel. Each knowledge group executes an inference by receiving an input value, and outputs the result. The inference result is led to a learning function activation determination unit. The learning function activation determining means evaluates the necessity of learning by comparing the inference result with a valid value. The legitimate value represents the actual result for the input value.
For example, in a stock price prediction system, various parameters for prediction are input values, and a valid value is an actual stock price. The learning function activation determining means further evaluates the necessity of the learning, and then starts and ends the learning according to the evaluation result.
End timing . The method of evaluating the necessity of learning is generally performed depending on whether the difference between the above inference result and the valid value is large or not. In general, the length of the learning period also changes according to the magnitude of the difference.

The learning is started by the learning function activation determining means.
When the learning period between the end timing and the end timing is set, the knowledge group inference unit updating means is activated in that period, the knowledge group is learned from the input value and the valid value, and the knowledge group in the knowledge group inference means is updated. Update.

In updating the knowledge group by the knowledge group inference unit updating means, it is possible to modify the existing knowledge group. However, the learned knowledge group is created as a new knowledge group, and the new knowledge group is created. It is also possible to add it as a new knowledge group.

[0014]

[0015]

FIG. 1 is a block diagram of an inference apparatus according to an embodiment of the present invention.

The input value input device 1 inputs data for executing inference. The input values are, for example, various prediction parameters when applied to a stock price prediction system.

The knowledge group inference unit 2 includes a plurality of knowledge groups 2a to 2
c. Each of the knowledge groups 2a to 2c receives input values from the input value input device 1 at the same time. Each knowledge group includes a fuzzy calculation unit, and performs a fuzzy rule matching calculation for each knowledge, a calculation for obtaining an inference result for each knowledge based on the matching, and a calculation for synthesizing the inference result for each knowledge. In the present embodiment, the knowledge group inference unit 2 includes n knowledge groups, and outputs a maximum of n inference results.

The valid value input device 3 inputs a valid value corresponding to the input value. For example, when applied to a stock price prediction system, the legitimate value is the actual stock price. The input value input from the input value input device 1 and the valid value input from the valid value input device 3 have a relationship in which the valid value is input once after the input value. Therefore, in terms of timing, first, an input value is input, then an inference result is output, and then a valid value is input.

The learning function determination device 4 outputs a correct value for each knowledge group from the inference result output from each knowledge group of the knowledge group inference unit 2 and the valid value input from the valid value input device 3. Evaluate if As described later, a predetermined threshold value is used for evaluation. When it is evaluated that no knowledge group outputs a correct value, the knowledge group learning device 9 is started.

The weight adjustment device 5 adjusts the weight value of the inference result of each knowledge group from the evaluation value of the inference result of each knowledge group evaluated by the learning function activation determination device 4. in this case,
The weight value for a knowledge group that outputs a correct value increases, and the weight value for a knowledge group that does not output a correct value decreases.

The weighting devices 51 to 53 actually weight the inference results of each knowledge group based on the weight values obtained by the weight adjusting device 5.

The inference result synthesizing unit 6 includes the weighting unit 5
The weighted inference results of each knowledge group appearing in the outputs 1 to 53 are synthesized. The synthesized inference result output device 7 outputs the synthesized inference result obtained from the inference result synthesis device 6. In this case, examples of the output device include a file, a line, and a display device.

The output control device 8 prevents the inference result from being output to the outside when the learning function activation determination device 4 activates the learning to the knowledge group learning device 9. To the inference result output device 7 after synthesis.

The knowledge group learning device 9 includes the input value input device 1
A new knowledge group is created based on the input value input from the valid value input device 3 and the valid value input from the valid value input device 3.

The knowledge group adding device 10 adds a new knowledge group obtained by the knowledge group learning device 9 to the knowledge group inference unit 2.

The above-mentioned knowledge group learning device 9 and knowledge group addition device 1
A means for updating the knowledge group with 0 (knowledge group inference unit updating means) is configured.

Next, the learning function activation determination device will be described in detail.

FIG. 2 shows the operation of evaluating the necessity of learning from the inference result of each knowledge group and the valid value.

First, the difference value and the differential value are calculated. The difference value is a difference between the inference result and the valid value.

Ei = | PVi-RV | ei: Difference between inference result of i-th knowledge group and valid value PVi: Inference result of i-th knowledge group RV: valid value This is the difference between the difference value at the time and the difference value this time. The calculation of the differential value is performed by calculating the difference between the difference value obtained in the calculation of the difference value and the difference value calculated in the previous inference.

E'i = | ei-ei (old) | e'i: Differential value of i-th difference value ei (old): Difference value of i-th previous inference value After the above operation, ei Is input to ei (old).

Subsequently, comparison with a threshold value is performed. If both the difference value and the derivative value exceed the threshold value, 1 is set in the learning function activation determination table. If the threshold is not exceeded, 0 is set. The threshold values of the difference value and the differential value are set in advance (for example, 5
%). FIG. 3 shows a learning function activation determination table.
It consists of an m × n two-dimensional array, and stores inference results for the past m times for each of n knowledge groups. Previous
The initial value of the table may be any value. m
After that time it will work properly. The values in the table are shifted left by one inference. For example, in the knowledge group 1, since the difference value and the differential value of the current inference both exceed the threshold value, 1 is set in the corresponding flag. If the flag is set to 1, it indicates that the inferred value is far from the legal value. After performing the above test for all knowledge groups, the flags are totaled. FIG. 4 shows a tabulation table when evaluating (monitoring) the inference results of the past 10 times. Aggregation (addition) is performed in the row direction of the aggregation table, and the aggregation result is written on the right end. The example illustrated in FIG. 4 indicates that the inference results of the past nine times of the knowledge group 1 match the legal values, and indicates that the current inference value is far from the legal values. Further, all of the knowledge groups 2 to n indicate that the inference value is far from the legal value.

Subsequently, the tally of the flags is compared with a threshold value. FIG. 4 shows an example of an aggregation table in the case of evaluating (monitoring) the inference results of the past 10 times. The totaling result of each knowledge group is compared with a threshold. If the total of the flags of all the knowledge groups exceeds the threshold, the learning function is activated. If one or more of the knowledge groups does not exceed the threshold, normal processing is performed. In this case, the weight adjustment device 5 in FIG. 1 increases the weight for the output of the knowledge group that does not exceed the threshold. Assuming that the threshold value is 7 in the example shown in FIG. 5, the total of the flags of all the knowledge groups exceeds the threshold value. This indicates that the inference results of all the knowledge groups are far from the legal values. In other words, it indicates that it is impossible to infer a valid value for the current knowledge group. In this case, learning is started. Note that, as learning data, in addition to the input value and the valid value used in the past ten inferences, the input value and the valid value until the next learning is completed are used.

The reason why the difference value and the differential value are used to determine whether the inferred value matches the legal value is to avoid erroneous determination when the respective values match by chance.

FIG. 6 shows the operation of the learning function activation determination device 4 after the start of learning.

The operation shown in FIG. 2 is the same as the operation up to the flag totaling step. As a result of the tallying of the flags, it is determined whether or not the tallying value of the flag of a certain knowledge group is equal to or smaller than a threshold value. If not, the learning continues. FIG. 7 shows an example of a tabulation table when evaluating (monitoring) the inference results for ten times. In this example, assuming that the threshold value is 3, the total value of the flags of the knowledge group 2 is equal to or smaller than the threshold value. This indicates that the inference result of the knowledge group 2 substantially matches the legal value. That is, it indicates that the knowledge group 2 can infer a valid value. In this case, the learning ends.

With the above operation, the learning function activation determining device 4 receives the inference result and the valid value of each knowledge group, evaluates the necessity of learning for updating the knowledge group, and uses the threshold value. This automatically sets the learning start timing and the learning end timing. Knowledge group inference unit The knowledge group learning device 9 and the knowledge group adding device 10 that constitute the updating means operate as follows.

That is, when receiving the learning start signal from the learning function activation determining device 4, the knowledge group learning device 9
Learning is performed using the input value and the valid value used during the past ten inferences and the valid value until the inference result and the valid value match next, that is, until the learning end signal is generated, and a knowledge group is generated. . The knowledge group adding device 10 adds the knowledge generated by the knowledge group learning device 9 to a free area of the knowledge group inference unit 2. FIG. 8 shows the configuration of the knowledge group inference unit 2. The knowledge group inference unit has a function of inferring N knowledge groups. Normally, knowledge groups are not stored in all N, and n are used for inference.

The inference units in which no knowledge groups are stored are referred to as knowledge groups a and b. If the knowledge group is not stored, the flag 11 is set to 1. If the knowledge group is stored, 0 is set in the flag 11. In the example shown in FIG. 8, since the flag 1 is set for the knowledge groups a and b, it indicates that no knowledge is stored in these knowledge groups a and b. In this state, when a new knowledge group is added by the knowledge group adding device 10, the new knowledge group is buried in order from the knowledge group a, and the flag 11 becomes 0. FIG.
Indicates the operation of the knowledge group adding apparatus 10. When a new knowledge group is generated by the knowledge group learning device 9, the state of the flag 11 is checked in order from the top by using a pointer. When a flag in which no knowledge group is stored is detected, the newly generated knowledge group is stored in the knowledge group storage area, and the corresponding flag is set.

In the above embodiment, when a new knowledge group is generated, the new knowledge group is sequentially added to the empty area. Since it is possible to judge the knowledge group that has exceeded, it is also possible to delete such a knowledge group from the inference unit 2 and increase the free area. Instead of generating and adding a new knowledge group, it is also possible to call an appropriate knowledge group and correct the knowledge group as a result of learning.
The above operation is summarized as follows. Input value input device
When data for executing inference is input from the device 1,
The knowledge group inference unit 2 receives the input data as an input value.
You. At this time, the knowledge group learning device 9 also
As input values. The knowledge that received the input value
The group inference unit 2 performs fuzzy inference and infers for each knowledge group.
Output the result. After outputting the inference result, is the valid value input device 3
The valid value is input from the
Each knowledge group outputs a correct value based on the theoretical result and the valid value
It evaluates whether or not the evaluation is performed, and stores the evaluation result. Said
Evaluation of the inference result evaluated by the learning function activation determination device 4
Based on the value, the inference result is obtained by the weight adjustment device 5.
Are adjusted by weighting devices 51-53.
Is weighted to the above inference result. Weighting of each knowledge group
The inference result is synthesized by the inference result synthesis device 6, and the synthesized inference is performed.
It is output to the logical result output device 7. On the other hand,
The determination device 4 determines the evaluation result of each knowledge group and the
Judge the necessity of learning by summing up the evaluation values of the last 10 times
You. When the learning function activation determination device 4 needs learning
When the determination is made, the knowledge group learning device 9 is activated to perform learning.
To start. At this time, the output control device 8
The synthesized inference result output device so as not to output the inference result
7 for that purpose. The knowledge group that started learning
The learning device 9 uses the input values used during the past ten inferences and the positive values.
In addition to this value, the knowledge group learning device 9 receives the
An input value and the correct value input from the correct value input device 3.
Learn using this value. Learning with the knowledge group learning device 9
Is generated by the knowledge group adding device 10.
Added to Knowledge Group Inference Unit 2 and updated Knowledge Group Inference Unit 2
You. After updating the knowledge group inference unit 2, the input value input device 1
Then, the following data is input. And the same flow as above
The learning function activation determination device 4 terminates the learning.
This is repeated until it is determined to be complete.

[0041]

According to the present invention, it is possible to recognize a situation which cannot be handled by the knowledge group of the current state knowledge group inference means, and to instruct the knowledge group update means to start and end a new learning in accordance with a change in the state of the inference target. .

Data in a situation that cannot be handled by any knowledge group can be collected as learning data, and necessary and sufficient learning data can be provided to the knowledge group updating means, so that the knowledge group updating means can be used effectively.

As a result of learning, a newly created knowledge group is added to the knowledge group inference means, and the situation in which the knowledge group inference device cannot cope can be reduced.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an embodiment of the present invention.

FIG. 2 is a flowchart illustrating a learning start determination operation of the learning function activation determination device.

FIG. 3

FIG. 5 shows a learning function activation determination table (total table).

FIG. 6 is a flowchart illustrating a learning end determination operation performed by the learning function activation determination device.

FIG. 7 is a diagram showing an aggregation table.

FIG. 8 is a diagram showing a knowledge group inference unit and flags connected to each knowledge group.

FIG. 9 is a flowchart showing the operation of the knowledge group adding device.

[Explanation of symbols]

 1-input value input device 2-knowledge group inference unit 2a-2c-knowledge group 3-proper value input device 4-learning function activation determination device 9-knowledge group learning device 10-knowledge addition device

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-3-67358 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06F 9/44 G06N 3/00-7 / 06 G05B 13/02

Claims (2)

(57) [Claims] 1. A knowledge group inference means for receiving an input value and performing inference, receiving an inference result and a valid value of the knowledge group inference means, evaluating necessity of learning for updating a knowledge group, and starting learning. When
Learning function activation determining means for determining an end timing; and learning function activation determining means determining that learning is necessary.
Comprising and a knowledge group inference unit updating means for updating the knowledge groups within the knowledge group inference means to learn knowledge group from the input value and the legal value when, with learning function inference apparatus odor
In addition, the learning function activation determining means may be configured to provide the information for each inference cycle.
Compare the inference result of each knowledge group of the knowledge group inference means with its legal value.
The stored results are stored as the evaluation value of each knowledge group.
Evaluation values for the past m inference cycles (m> 1)
The necessity of learning is determined for each inference cycle based on
An inference apparatus with a learning function, which determines the start and end timings of the learning. 2. The knowledge group inference unit updating means includes: knowledge group learning means for creating a learned knowledge group as a new knowledge group;
The inference device with a learning function according to claim 1, further comprising: a knowledge group adding unit that adds a new knowledge group to the knowledge group inference unit.