Learning device

Abstract

To provide a learning device that can quickly update mapping data with the latest values of input variable values, compared to a case where a dataset is prepared.SOLUTION: A PU 12 calculates a predicted value for a power generation amount by a wave power plant after a prescribed period of time by inputting time-series data of wave height values measured by a wave height meter 20 into a regression model. The PU 12 updates mapping data 14b that defines the regression model so as to reduce a difference between the power generation amount obtained after the prescribed period of time and the predicted value. The PU 12 repeats the above process of updating the mapping data 14b each time time-series data of wave height values is obtained, without preparing a dataset.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a learning device. [Background technology]

[0002] For example, as described in Non-Patent Document 1 below, machine learning involves dividing a dataset into a training set, a validation set, and a test set for learning. The training set is used to update parameters, for example, in the case of learning a parametric model. Parameter updating using the training set is performed by batch processing, mini-batch processing, or online learning. Cross-validation may also be used to ensure a sufficient training set for a limited dataset.

[0003] If predictions, etc. are continued to be performed using a trained model that has been trained using the above dataset, there is a risk that the prediction accuracy will decrease due to changes in the situation. Therefore, it has been proposed to retrain a trained model using a new data set at a predetermined update timing, as seen in, for example, Patent Document 1 below. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2023-118355 [Non-patent literature]

[0005] [Non-Patent Document 1] C.M. Bishop, "Pattern Recognition and Machine Learning," Maruzen Publishing, January 20, 2012, pp. 31-33 Summary of the Invention [Problem to be solved by the invention]

[0006] However, when learning using a dataset as described above, it is necessary to first prepare the dataset each time a request arises to generate or update a trained model. [Means for solving the problem]

[0007] The means for solving the above problems and their effects will be described below. 1. A learning device comprising: a storage device and an execution device, wherein mapping data is stored in the storage device, the mapping data being data for defining a mapping that outputs a value of an output variable in response to input of a value of an input variable, the values ​​of the input variables including values ​​that newly appear over time, and the output variables being variables that indicate the result of inference by the mapping, the execution device being configured to execute an acquisition process and an update process, wherein the acquisition process is a process of acquiring values ​​of the input variables and target values ​​for the output variables, and the update process is a process of updating the mapping data so as to reduce an error between the value of the output variable and the target value when the value of the input variable is input to the mapping, and wherein the process is triggered by the acquisition process of new values ​​of the input variables and the target value, and is executed repeatedly without making the values ​​of the input variables and the target value into a data set.

[0008] Since the values ​​of the input variables are updated over time, the trigger for executing the update process is generated repeatedly over time. Therefore, in the above configuration, the update process is repeatedly executed each time a new value of the input variable is acquired. Therefore, compared to the case where a data set is prepared, the mapping data can be updated more quickly with the latest values ​​of the input variables.

[0009] 2. The learning device according to 1 above, which is configured to remove the values ​​of the input variables once input to the mapping for the update process from the training data used to update the mapping data. In the past, a dataset consisting of a set of input variable values ​​and target values ​​was divided into a training set, a validation set, and a test set, and then the mapping data was updated by repeatedly using the training set. In other words, the values ​​of input variables that were once used to calculate output variables were repeatedly used to calculate output variable values. In this case, the computational load required for learning increased and the learning time became longer.

[0010] In contrast, with the above configuration, the values ​​of input variables once input to the mapping are removed from the training data. This reduces the computational load required for learning and shortens the learning time. Therefore, the above configuration is particularly suitable for situations where the values ​​of input variables are updated sequentially over time.

[0011] 3. A learning device as described in 1 or 2 above, wherein the execution device is configured to execute a providing process, which is executed when there is a history of the error being below a predetermined value, and which is a process of providing the inference result by the mapping to the outside depending on the value of the output variable.

[0012] In the above configuration, when the accuracy of the output variables of the mapping defined by the mapping data, which is updated each time a new input variable value is acquired, increases, the inference results are provided to the outside, thereby reducing the number of steps from the start of learning to the start of providing the inference results.

[0013] In contrast, in the past, a sufficient number of pairs of input variable values ​​and target values ​​were collected to create a dataset, and then the dataset was divided into a training set, a validation set, and a test set. After updating the mapping data using the training set, validation using the validation set and testing using the test set were performed, resulting in a trained model whose inference results can be provided to external parties. This resulted in a large number of steps from the start of training to the start of providing inference results.

[0014] 4. A learning device as described in claim 3, wherein the update process includes a process of updating the mapping data used to calculate the value of the output variable using the value of the output variable used in the provision process in order to reduce the error in the value of the output variable used in the provision process.

[0015] Conventionally, when performing learning and inference simultaneously, it was necessary to prepare two types of models: a model to be learned and a deployed inference model. In such cases, the re-learned content could not be reflected in the inference model until re-learning was completed. Furthermore, even if new input variable values ​​and target values ​​were acquired through an acquisition process during learning, these newly acquired values ​​could not be used in ongoing learning. In other words, because learning is performed after a dataset consisting of multiple pairs of input variable values ​​and target values ​​is divided into a training set, a validation set, and a test set, newly acquired data during learning could not be used until a new re-learning opportunity was offered after learning was completed.

[0016] In contrast, the above configuration allows sequential learning using the values ​​of the output variables used for inference, so the mapping data can be quickly updated with the latest combination of input variable values ​​and target values. Also, by sharing the process of inputting input variable values ​​into the mapping and calculating output variable values ​​between inference and learning, the total task load on the hardware resources can be reduced.

[0017] 5. A learning device according to any one of claims 1 to 4, configured to execute an environmental information acquisition process, wherein the environmental information acquisition process is a process of acquiring environmental information, and the update process includes a process of changing a learning rate according to the environmental information acquired by the environmental information acquisition process.

[0018] The mapping defined by the mapping data is affected by the environment, which is expressed by variables that are not directly included in the input variables. If inference based on the mapping is expected to be used in a specific environment, it is desirable for the learning rate in that specific environment to be higher than the learning rate in other environments. Furthermore, if the environment changes significantly, the mapping defined by the mapping data learned before the change may no longer be appropriate for the changed environment. Therefore, it is desirable to change the learning rate depending on whether inference in the environment before or after the change is emphasized.

[0019] Therefore, in the above configuration, by changing the learning rate in accordance with the environmental information, it becomes possible to appropriately update the mapping data in accordance with the environment in which the inference results are to be used. 6. A learning device as described in 5 above, configured to execute an environmental history storage process, wherein the environmental history storage process is a process for storing a history of environmental information acquired by the environmental information acquisition process, and the update process includes a process for changing a learning rate when the environmental information acquired by the environmental information acquisition process deviates from the stored environmental information by a predetermined amount or more.

[0020] The mapping defined by the mapping data before the environment changed significantly may not be an appropriate mapping for the changed environment. Also, when the environmental change is temporary, it may be desirable to utilize the inference performance of the mapping learned before the change, but if the mapping data is updated due to the environmental change, there is a risk that the inference accuracy of the mapping will decrease when the environment returns to its original state.

[0021] Therefore, in the above configuration, when the environment changes significantly, by changing the learning rate, it becomes possible to appropriately update the mapping data according to the environment in which the inference results are to be used. 7. A learning device as described in 4 to 6 above (excluding those not dependent on 4 above) in which the update process includes a process of changing the learning rate in accordance with the error, under the condition that the learning rate when the error in the value of the output variable used in the providing process relative to the target value is large is equal to or greater than the learning rate when the error is small.

[0022] In the above configuration, by increasing the learning rate when the error is large, the mapping data can be updated quickly to reduce the error. 8. A learning device as set forth in any one of claims 3 to 7 (excluding those not dependent on claim 3) that does not execute the update process if the error in the value of the output variable used in the provision process relative to the target value is less than a predetermined value.

[0023] In the above configuration, the update process is not executed when the error is equal to or less than a predetermined value, thereby making it possible to prevent overlearning from occurring. [Brief explanation of the drawings]

[0024] [Figure 1] FIG. 1 illustrates a learning device according to an embodiment. [Figure 2] 2 is a flowchart showing the procedure of processing executed by the learning device of FIG. 1. [Figure 3] 2 is a flowchart showing the procedure of processing executed by the learning device of FIG. 1. DETAILED DESCRIPTION OF THE INVENTION

[0025] Hereinafter, an embodiment will be described with reference to the drawings. "Prerequisite configuration" FIG. 1 shows the configuration of a model learning system in this embodiment.

[0026] The learning device 10 includes a PU 12, a storage device 14, and a communication device 16. The PU 12 is a software processing device such as a CPU, a GPU, or a TPU. The storage device 14 may be an electrically non-rewritable non-volatile memory. Alternatively, the storage device 14 may be an electrically rewritable non-volatile memory or a storage medium such as a disk medium. A learning program 14a is stored in the storage device 14.

[0027] The learning device 10 executes a process of generating a regression model by having the PU 12 execute the learning program 14a. The regression model according to this embodiment is a mapping in which the wave height values ​​WH1, WH2, ... detected by the multiple wave height meters 20 are input, and the predicted value APe of the power generation amount AP of the wave power plant after a predetermined period of time is output. The multiple wave height meters 20 are arranged at different positions around the wave power plant.

[0028] The regression model may be, for example, a parametric model. Specifically, it may be, for example, a neural network. Alternatively, the regression model may be a non-parametric model. Specifically, it may be, for example, support vector regression.

[0029] The storage device 14 stores mapping data 14b, which is data for defining a regression model. When the regression model is a neural network, the mapping data 14b includes data indicating the weighting coefficient values ​​and bias values ​​of the neural network. When the regression model is support vector regression, the mapping data 14b includes data indicating support vectors.

[0030] The learning device 10 executes a process of displaying the predicted value APe based on the regression model on the display unit 24. "Learning Device Processing" 2 and 3 show the procedure of the learning process performed by the learning device 10. The process shown in Fig. 2 and 3 is realized by the PU 12 repeatedly executing the learning program 14a, for example, at a predetermined interval. Note that, below, the step number of each process is represented by a number preceded by "S."

[0031] In the series of processes shown in FIG. 2, the PU 12 first acquires weather information via the communication device 16 (S10). The acquired weather information includes, for example, temperature, wind speed, precipitation, and snowfall. Next, the PU 12 acquires time-series data of wave height values ​​WH1, WH2, WH3, ... (S12). The wave height values ​​WH1, WH2, WH3, ... are values ​​detected by each of the multiple wave height meters 20. In FIG. 2, the numbers in parentheses after each of the wave height values ​​WH1, WH2, WH3, ... identify the sampling numbers. For example, wave height value WH1(2) means that the sampling was performed after wave height value WH1(1).

[0032] Next, the PU 12 calculates the predicted value APe of the amount of power generation by inputting the time series data acquired by the process of S10 into a regression model defined by the mapping data 14b (S14).

[0033] Then, the PU 12 temporarily stores the weather information acquired in the process of S10, the time-series data acquired in the process of S12, and the predicted value APe acquired in the process of S14 in the storage device 14 (S16).

[0034] Next, the PU 12 determines whether the flag F is "1" (S18). The value of the flag F being "1" indicates that the predicted value APe based on the regression model is available. The value of the flag F being "0" indicates that the predicted value APe based on the regression model is not yet available. The value of the flag F is set to "0" when the mapping data 14b is first updated.

[0035] If the determination in the process of S18 is affirmative, the PU 62 outputs the predicted value APe to the display unit 24 (S20). Since the process of S20 causes the predicted value APe to be displayed on the display unit 24, a person can use the output of the regression model.

[0036] When the process of S20 is completed or when a negative determination is made in the process of S18, the PU 12 proceeds to the process of S22 shown in Fig. 3. In the process of S22, the PU 12 acquires the power generation amount AP detected by the power meter 22 shown in Fig. 1.

[0037] Next, the PU 12 determines whether the flag F is "1" (S24). If the PU 12 determines that the flag F is "0" (S24: NO), the PU 12 determines whether the condition (A) that the absolute value of the difference between the power generation amount AP and the predicted value APe is equal to or smaller than the threshold value Δth has continued for a predetermined number of times has been satisfied (S26). This process is a process for determining whether the predicted value APe has become available as the learning of the regression model progresses.

[0038] When it is determined that the above condition (A) is satisfied (S26: YES), the PU 62 assigns "1" to the flag F (S28). That is, the satisfaction of the condition (A) in this embodiment corresponds to the selection of a model to be deployed after undergoing testing using a test set of a dataset in conventional machine learning.

[0039] When the PU 12 completes the process of S28, when the PU 12 makes a positive determination in the process of S24, or when the PU 12 makes a negative determination in the process of S26, the PU 12 reads out data from a predetermined period of time ago from the data temporarily stored in the process of S16 (S30). The predetermined period of time indicates how far into the future the amount of power generation indicated by the output of the regression model is from the present.

[0040] The PU 12 determines whether the weather information read by the process of S30 deviates from the weather information history at the time of acquisition of the time-series data of wave height values ​​WH1, WH2, WH3, ... used to update the mapping data 14b (S32). Here, the weather information history data 14c stored in the storage device 14 by the process of S44 described later is referenced. The PU 12 determines that the weather information deviates, for example, when the wind speed indicated by the weather information read by the process of S30 is greater than the maximum wind speed indicated by the history data 14c by a predetermined speed or more.

[0041] When it is determined that the weather information is deviating (S32: YES), the PU 12 sets the learning rate α of the regression model to a high rate αH (S34). On the other hand, when it is determined that the weather information does not diverge (S32: NO), the PU 12 determines whether the following condition (B) is met (S36).

[0042] Condition (B): The absolute value of the difference between the predicted value APe read in the process of S30 and the power generation amount AP obtained in the process of S22 is equal to or greater than a threshold value ΔthH for a predetermined number of consecutive times. This condition indicates that the accuracy of the regression model is still quite low. The threshold value ΔthH is set to a value greater than the threshold value Δth.

[0043] If the PU 12 determines that the condition (B) is met (S36: YES), the PU 12 proceeds to the process of S34. On the other hand, if the PU 12 determines that the condition (B) is not met (S36: NO), the PU 12 determines whether the following condition (C) is met (S38).

[0044] Condition (C): A state in which the absolute value of the difference between the predicted value APe read by the process of S30 and the power generation amount AP obtained by the process of S22 is equal to or greater than a threshold value ΔthL continues for a predetermined number of times. This condition indicates that the regression model has been sufficiently trained. The threshold value ΔthL is set to a value smaller than the threshold value ΔthH. The threshold value ΔthL may be set to a value smaller than the threshold value Δth.

[0045] When the PU 12 determines that the condition (C) is not satisfied (S38: NO), it sets the learning rate α to a low rate αL (S40). When the processes of S34 and S40 are completed, the PU 12 updates the mapping data 14b in accordance with the learning rate α so as to reduce the absolute value of the difference between the predicted value APe read in the process of S30 and the power generation amount AP obtained in the process of S22 (S42). For example, if the regression model is a neural network, the PU 12 may update the weighting coefficients and bias values ​​indicated by the mapping data 14b by the backpropagation method in accordance with the difference.

[0046] Then, the PU 12 adds the data indicating the weather information read in the process of S30 to the history data 14c stored in the storage device 14 shown in FIG. 1 (S44). When the process of S44 is completed or when a positive determination is made in the process of S38, the PU 12 temporarily ends the series of processes shown in FIGS.

[0047] "Actions and Effects of the Present Embodiment" Each time the PU 12 acquires new time-series data of peak values ​​WH1, WH2, WH3, ..., it inputs the data as training data into the regression model to calculate the predicted value APe. Each time a predicted value APe is calculated, the PU 12 updates the mapping data 14b so that the difference between the predicted value APe and the corresponding power generation amount AP becomes smaller. As described above, in this embodiment, no data set is prepared for updating the mapping data 14b. This reduces the number of steps required to update the mapping data 14b, and enables the mapping data 14b to be updated quickly.

[0048] According to the present embodiment described above, the following actions and effects can be further obtained. (1) After the update process of the mapping data 14b starts, if the condition (A) is satisfied, the PU 12 determines that the regression model has been trained and makes the predicted value APe available. In other words, after the start of learning of the regression model, if the condition (A) is satisfied, the regression model has been trained and makes the predicted value APe available. This makes it possible to determine the completion of learning without the steps of preparing a data set of peak values ​​WH1, WH2, WH3, ... and power generation amount AP and dividing it into a training set, a validation set, and a test set.

[0049] (2) Even when the PU 12 provides the predicted value APe as available data, the PU 12 updates the mapping data 14b so that the difference between the predicted value APe and the power generation amount AP obtained after a predetermined period becomes small. This allows learning and inference to be performed simultaneously.

[0050] Previously, when training and inference were performed simultaneously, two models were required: a training model and a deployed inference model. In this case, the retrained data could not be reflected in the inference model until retraining was completed. Furthermore, even if new samples of peak values ​​WH1, WH2, WH3, ... and power generation amount AP were taken during training, these sampled values ​​could not be used in ongoing training. In other words, previously, training was performed after the dataset of peak values ​​WH1, WH2, WH3, ... and power generation amount AP was divided into a training set, a validation set, and a test set. Therefore, the newly sampled data during training could not be used until a new retraining opportunity was available after training was completed.

[0051] In contrast, in this embodiment, since learning and inference can be performed simultaneously, the accuracy of inference can be quickly improved using the latest sampled values. Furthermore, the process of calculating the predicted value APe based on the peak values ​​WH1, WH2, WH3, ... as input variables is shared between learning and inference, thereby reducing the total number of tasks required by hardware resources.

[0052] (3) The PU 12 removes data that has been used to update the mapping data 14b from the data used to update the mapping data 14b. This is effective in reducing the amount of calculation required for learning when new data is generated each time, and in quickly reflecting each new data in the mapping data 14b.

[0053] In other words, in the past, data that had been input once into the regression model was input multiple times for batch processing using a training set of data sets of peak values ​​WH1, WH2, WH3, ... and power generation amount AP, which increased the amount of calculation required for learning and lengthened the time required for re-learning.

[0054] <Correspondence> The correspondence between the matters in the above embodiment and the matters described in the "Means for Solving the Problem" column is as follows. Below, the correspondence is shown for each number of the means for solving the problem described in the "Means for Solving the Problem" column. [1] The storage device corresponds to the storage device 14. The execution device corresponds to the PU 12. The detected values ​​of the sensors correspond to the peak values ​​WH1(1), WH1(2), ..., WH2(1), WH2(2), .... The output variable corresponds to the predicted value APe of the power generation amount. The target value corresponds to the power generation amount AP. The acquisition process corresponds to the processes of S12 and S22. The update process corresponds to the process of S42. [2] The process of S42 is performed only once using the data newly acquired by the processes of S12 and S22. [3] The provision process corresponds to the process of S20. [4] The predicted value APe used in the process of S20 becomes the input for the process of S42. [5, 6] The environmental information acquisition process corresponds to the process of S10. The environmental history storage process corresponds to the process of S44. The process of changing the learning rate corresponds to the fact that the learning rate may be larger when the determination in the process of S32 is affirmative than when the determination in the process of S32 is negative. [7] This corresponds to the fact that the learning rate may be larger when the determination in the process of S36 is affirmative than when the determination in the process of S36 is negative. [8] This corresponds to the fact that the process of S42 is not executed when the determination in the process of S38 is affirmative.

[0055] <Other embodiments> This embodiment can be modified as follows: This embodiment and the following modifications can be combined with each other within the scope of technical compatibility.

[0056] "About learning rates" The learning rate setting process is not limited to a process of setting it to one of two different values. For example, the learning rate setting process may be a process of setting it to three different values, such as setting the base learning rate when flag F is zero to a value greater than the low rate αL and less than the high rate αH. Furthermore, the learning rate setting process may be a process of setting it to four or more different values. Here, the learning rate may be changed depending on the error, provided that the learning rate when the error is large is equal to or greater than the learning rate when the error is small.

[0057] In the description "changing B according to A under the condition that when A is large, B is equal to or larger than B when A is small," the case where A is large and the case where A is small refer to the relative relationship of magnitude when comparing the two. For example, "when A is large" corresponds to the case where "A is a first value," and "when A is small" corresponds to the case where "A is a second value smaller than the first value." The above description also means that, depending on the settings of the first and second values, B when A is the first value may be larger than B when A is the second value. The above description also means that B is changed according to A so that A when B is large is larger than A when B is small.

[0058] "About Mapping" The mapping in which the predicted value APe is the output variable is not limited to the mapping exemplified in the above embodiment. For example, the input variables may include a variable indicating wind speed and a variable indicating wind direction.

[0059] · Regression models are not limited to models that predict power generation. The mapping does not necessarily have to be a regression model. It can also be a discriminative model. In that case, the magnitude of the error can be quantified using, for example, cross entropy.

[0060] The values ​​of the input variables of the model defined by the mapping do not necessarily include sensor detection values ​​that are updated over time. For example, the input variables may include each word in tweets containing a specific political term and its frequency of appearance. In this case, the output variable of the model may be the support rate for each political party. Since new values ​​emerge for the input variables of this model over time, it is effective to perform learning and inference in the same manner as in the above embodiment. For example, the input variable may be the number of accesses to an Internet homepage, and the output variable may be the average load rate of each of multiple servers and the total uptime of all servers. In this example, the average load rate may be replaced with the maximum load rate. In this example, the average load rate may be replaced with the percentage of load rates exceeding a certain threshold. In this example, the average load rate may be replaced with the variance and mean of a normal distribution, assuming that the load rate follows a normal distribution.

[0061] "About environmental information acquisition processing" Environmental information is not limited to weather information. For example, if a regression model outputs a predicted value for the sales of a certain store, information about surrounding stores may be used as environmental information. In this case, for example, if a new store opens, the learning rate may be increased because the likelihood of sales changing increases.

[0062] "About the update process" If the environmental information significantly deviates from the past environmental information when the update process was executed, it is not necessary to increase the learning rate in the update process. For example, if the environmental information significantly deviates from the past environmental information when the update process was executed, the learning rate in the update process may be reduced. Specifically, for example, the PU 12 may reduce the learning rate when an exceptional event such as a typhoon or tsunami occurs.

[0063] "About learning devices" The learning device that performs inference and learning is not limited to a single PU 12. For example, inference and learning may be performed by parallel calculations using multiple PUs. Furthermore, for example, one of the multiple PUs with a small computational load each time may perform inference and learning. It is not essential that the PU that performs inference and the PU that performs learning are the same. In other words, it is not essential that the process of calculating the value of an output variable by inputting the value of an input variable into a mapping be shared between inference and learning; the PU that performs inference and the PU that performs learning may each perform the process.

[0064] The learning device is not limited to one that executes software processing. For example, it may be equipped with a dedicated hardware circuit, such as an ASIC, that executes at least part of the processing executed in the above embodiments. That is, the learning device may include a processing circuit having any of the following configurations (a) to (c): (a) A processing circuit that includes a processing device that executes all of the above processing in accordance with a program, and a program storage device, such as a memory device, that stores the program. (b) A processing circuit that includes a processing device and program storage device that executes part of the above processing in accordance with a program, and a dedicated hardware circuit that executes the remaining processing. (c) A processing circuit that includes a dedicated hardware circuit that executes all of the above processing. Here, there may be multiple software execution devices that include a processing device and a program storage device. Also, there may be multiple dedicated hardware circuits. [Explanation of symbols]

[0065] 10...Learning device 14...Storage device 14a…Study Program 14b...Mapping data 14c...Historical data 16...Communication device 20...Wave height gauge 22…Power meter 24…Display section

Claims

1. a storage device and an execution device; The storage device stores mapping data, the mapping data is data for defining a mapping that outputs a value of an output variable in response to an input of a value of an input variable, the values ​​of the input variables include newly emerging values ​​over time; the output variable is a variable indicating a result of the inference by the mapping, the execution device is configured to execute an acquisition process and an update process; the acquisition process is a process of acquiring values ​​of the input variables and target values ​​for the output variables, The update process is a process of updating the mapping data so as to reduce the error of the value of the output variable relative to the target value when the value of the input variable is input to the mapping, and is a process that is repeatedly executed without making the value of the input variable and the target value into a data set, triggered by the acquisition process of new values ​​of the input variables and the target value.

2. 2. The learning device according to claim 1, wherein the values ​​of the input variables once input to the mapping for the update process are removed from the training data used to update the mapping data.

3. the execution device is configured to execute a provisioning process; The learning device according to claim 1, wherein the provision process is executed when there is a history of the error being below a predetermined value, and is a process of providing the inference result by the mapping to an external device depending on the value of the output variable.

4. 4. The learning device according to claim 3, wherein the update process includes a process of updating the mapping data used to calculate the value of the output variable using the value of the output variable used in the provision process in order to reduce the error in the value of the output variable used in the provision process.

5. configured to execute an environmental information acquisition process; the environmental information acquisition process is a process of acquiring environmental information, The learning device according to claim 1 , wherein the update process includes a process of changing a learning rate in accordance with the environmental information acquired by the environmental information acquisition process.

6. configured to perform an environmental history storage process; the environmental history storage process is a process of storing a history of environmental information acquired by the environmental information acquisition process, The learning device according to claim 5 , wherein the update process includes a process of changing a learning rate when the environmental information acquired by the environmental information acquisition process deviates from the stored environmental information by a predetermined amount or more.

7. 5. The learning device according to claim 4, wherein the update process includes a process of changing the learning rate in accordance with the error, under the condition that the learning rate when the error in the value of the output variable used in the providing process relative to the target value is large is equal to or greater than the learning rate when the error is small.

8. The learning device according to claim 3 , wherein the update process is not executed when the error of the value of the output variable used in the providing process from the target value is equal to or less than a predetermined value.

Images