JP7208503B2 - Machine learning program, machine learning method and machine learning apparatus - Google Patents

Description

The present invention relates to a machine learning program, a machine learning method and a machine learning apparatus.

Machine learning is sometimes performed as data analysis using a computer. In machine learning, a computer is fed training data representing multiple instances with known outcomes. A computer analyzes training data to generate a predictive model that generalizes the relationship between factors (sometimes called explanatory or independent variables) and outcomes (sometimes called objective or dependent variables). do. An unknown result can be predicted by using the generated prediction model.

Machine learning is sometimes used for crop yield prediction. For example, a prediction device has been proposed that predicts the optimal harvest date of agricultural products. The proposed prediction device collects teacher data including images of crops taken on different days before harvest and harvest dates when the crops were actually harvested. The prediction device generates a prediction model from the teacher data by machine learning, inputs an image of the target crop to the prediction model, and predicts the harvest date of the target crop.

JP 2018-169993 A

In machine learning, which makes it possible to predict the yield of agricultural crops, a prediction model is generated that uses the growth environment such as temperature and amount of sunlight as explanatory variables, and the required number of days from the reference date, such as the date of fruiting, to the harvest date as the objective variable. can be considered. However, crops have individual differences in that they grow at different rates even if they are grown under the same growing environment. In particular, some types of crops have large individual differences. In contrast, general machine learning generates a predictive model that calculates one expected value (most likely value) of the objective variable for one value of the explanatory variable. As a result, although the actual harvest dates vary, the prediction model will concentrate the harvest dates for many crops on a specific date, which may lead to prediction results that deviate from the actual situation. be.

In one aspect, an object of the present invention is to provide a machine learning program, a machine learning method, and a machine learning apparatus that generate a prediction model with improved accuracy in crop yield prediction.

In one aspect, a machine learning program is provided that causes a computer to perform the following processes. Training data including multiple records each of which is associated with information on the growing environment of sample crops and the required number of days from the reference date when a predetermined state was observed to the harvest date of sample crops, and multiple samples indicated by the multiple records Total data indicating the actual distribution of the number of harvests with respect to the harvest date is acquired for the crop set including the crop and other crops. A learning process that repeats the process of generating a prediction model that calculates the probability distribution of the number of days required from the training environment information, evaluating the error in the probability distribution calculated by the prediction model using the training data, and updating the prediction model. Start. In the middle of the learning process, combine multiple probability distributions calculated by the prediction model from information on the growing environment indicated by multiple records, calculate the predicted distribution of the number of harvests for the harvest date, and calculate the predicted distribution and the total number data. The timing to stop the learning process is determined based on the degree of similarity with the actual distribution.

Also, in one aspect, a computer-implemented machine learning method is provided. Also, in one aspect, a machine learning device having a storage unit and a processing unit is provided.

In one aspect, it is possible to generate a prediction model with improved accuracy in crop yield prediction.

It is a figure explaining the example of the machine-learning apparatus of 1st Embodiment. It is a figure which shows the example of the information processing system of 2nd Embodiment. It is a figure which shows the hardware example of a machine-learning apparatus. It is a figure which shows the example of the data flow of harvest prediction. It is a figure which shows the usage example of the prediction model which outputs an expected value. It is a figure which shows the usage example of the prediction model which outputs a probability distribution. It is a figure which shows the example of use of the prediction model of under-learning. It is a figure which shows the usage example of the prediction model which over-trained. It is a figure which shows the stop timing example of machine learning. FIG. 4 is a diagram illustrating an example of a machine learning data flow; 3 is a block diagram showing an example of functions of a machine learning device; FIG. It is a figure which shows the example of a table of meteorological data, sample data, and total data. It is a figure which shows the example of a training data table. 5 is a flow chart showing an example of machine learning procedure. 10 is a flowchart (continued) showing an example of the machine learning procedure; It is a flowchart which shows the procedure example of harvest prediction.

Hereinafter, this embodiment will be described with reference to the drawings.
[First embodiment]
A first embodiment will be described.

FIG. 1 is a diagram illustrating an example of a machine learning device according to the first embodiment.
The machine learning device 10 according to the first embodiment uses machine learning to generate a prediction model used for crop yield prediction. The machine learning device 10 is also called an information processing device or a computer. The machine learning device 10 may be a client device or a server device.

A machine learning device 10 has a storage unit 11 and a processing unit 12 . The storage unit 11 may be a volatile semiconductor memory such as a RAM (Random Access Memory), or may be a non-volatile storage such as an HDD (Hard Disk Drive) or flash memory. The processing unit 12 is, for example, a processor such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a DSP (Digital Signal Processor). However, the processing unit 12 may include electronic circuits for specific purposes such as ASICs (Application Specific Integrated Circuits) and FPGAs (Field Programmable Gate Arrays). The processor executes a program stored in a memory such as RAM (which may be the storage unit 11). A collection of multiple processors is sometimes called a "multiprocessor" or simply a "processor."

Storage unit 11 stores training data 13 and total number data 14 . The training data 13 and the total data 14 are historical data regarding harvested crops. The training data 13 and the total data 14 indicate the growth status and harvest status of past harvested crops such as crops harvested in the previous year. The crops indicated by the training data 13 and the crops indicated by the total data 14 were harvested in the same year. The training data 13 and the total data 14 may be single-year data representing agricultural crops for one year, or may be multi-year data in which agricultural crops for multiple years are mixed. Crops may include fruits. Fruits are the edible fruits of plants, such as vegetables and fruits, and are grown by farmers. Agricultural crops have individual differences in growth even if they are grown in the same growing environment, resulting in variations in harvestable days. The crops may be of a variety with large individual differences in growth, such as paprika.

Training data 13 includes a plurality of records for sample crops that are part of an entire harvested crop (crop set). A sample crop in a set of crops is a collection of individual detailed information about the growth status and harvesting status. Other crops in the crop collection did not collect individual detailed information. The ratio of sample crops to the crop set (sample ratio) may be about 0.01% to 0.3%. This is because it takes time and effort to collect detailed information.

A plurality of records included in the training data 13 each associate information on the growing environment of the sample crop with the required number of days. The information on the growing environment includes indices that are correlated with the growth of crops, such as temperature and amount of solar radiation. For example, the growing environment information includes the following average temperature and average solar radiation from the reference date to the harvest date. However, if a correlation with the growth of crops is recognized, the temperature and the amount of insolation before the reference date may be used, or the cumulative temperature and the amount of insolation may be used. The base date may differ depending on the sample crop. If the reference date changes, the growing environment information associated with the sample crop may also change accordingly. The required number of days is the number of days from the reference date when a predetermined state of the sample crop was observed to the harvest date when the sample crop was harvested. For example, the reference date is the date when it is observed that the plant bears fruit (fruit bearing date). However, the reference date may be the day when the plant reaches a predetermined state before bearing fruit, or the day when the sample crop reaches a predetermined state after bearing fruit. When harvest management is performed on a weekly basis, the unit of required days may be a week.

The total number data 14 indicates the actual distribution of the number of harvests (the number of harvested crops) with respect to the harvest date for the crop set including the sample crops and other crops indicated by the training data 13 . Each crop is detached from the plant and harvested on the day the farmer determines it is mature enough. Harvest dates vary due to differences in fruiting dates and individual differences in growth. When harvest management is performed on a weekly basis, the total number data 14 may indicate the actual distribution of the number of harvests for the week to which the harvest date belongs. The total data 14 is collected for shipping management, and requires less effort than the training data 13 to collect.

For example, one record included in the training data 13 is information that the required number of days from the date of bearing fruit to the date of harvest was 8 weeks for a sample crop grown under a specific average temperature and average amount of solar radiation. indicates Another record included in the training data 13 is information that the number of days required from the date of bearing fruit to the date of harvest was 7 weeks for the sample crop grown under another average temperature and average amount of solar radiation. indicates Total data 14 shows that out of 12,000 crops including sample crops and other crops, 3,700 were harvested in one week, 5,800 were harvested in the next week, and 5,800 were harvested in the next week. shows the information that 2,500 were harvested.

The processing unit 12 executes a learning process 15 to generate a prediction model 16 . Various machine learning algorithms such as genetic programming (GP), multiple regression analysis, and neural network (NN) can be used to generate the prediction model 16 . The prediction model 16 is a statistical model that receives information on the growing environment as an explanatory variable and outputs a probability distribution of required days as an objective variable. When the training data 13 indicates the required number of days in weeks, the prediction model 16 may output the probability distribution of the number of weeks. The prediction model 16 is learned so as to output probabilities for each of a plurality of required days instead of outputting only the required days with the highest probability (expected value of required days). For example, the forecast model 16 outputs a probability distribution of 30% for 7 weeks, 50% for 8 weeks, and 20% for 9 weeks for a particular average temperature and average solar radiation.

In the learning process 15 , the processing unit 12 repeats updating the prediction model 16 by evaluating the error of the probability distribution calculated by the prediction model 16 using the training data 13 . For example, the processing unit 12 inputs the training environment information indicated by each of the plurality of records included in the training data 13 to the prediction model 16, and the prediction model 16 outputs using the required number of days indicated by the record. Evaluate the error of the probability distribution that Then, for example, the processing unit 12 updates the coefficients included in the prediction model 16 so that the error becomes smaller. For neural networks, the weights of edges (synapses) between nodes are updated.

Here, the processing unit 12 controls the number of iterations of updating the prediction model 16 in the learning process 15 . When the number of iterations is small, the probability distribution output by the prediction model 16 has a large error with respect to the training data 13 and a low fitting accuracy to the training data 13 . As the number of iterations increases, the error of the probability distribution output by the prediction model 16 with respect to the training data 13 decreases step by step, and the accuracy of fitting to the training data 13 increases step by step.

However, the number of sample crops represented by the training data 13 is small with respect to the entire set of crops, and there are individual differences in the growth of the crops. Therefore, the samples of required days indicated by the training data 13 do not necessarily faithfully represent the true probability distribution for the entire crop set, and bias exists. Therefore, if the number of iterations is increased too much, the prediction model 16 will fit the training data 13 excessively due to over-learning. The probability distribution output by the over-learned prediction model 16 has an excessively small variance, and may deviate from the true probability distribution in which the number of required days varies due to individual differences.

Therefore, the processing unit 12 refers to the total number data 14 and stops updating the prediction model 16 in the learning process 15 at an appropriate timing.
Specifically, in the middle of the learning process 15, the processing unit 12 calculates a plurality of probability distributions using the current prediction model 16 from the training environment information indicated by the plurality of records included in the training data 13. are combined to calculate a prediction distribution 17. The prediction distribution 17 is calculated, for example, each time the prediction model 16 is updated. The predicted distribution 17 is a predicted distribution of the number of harvests with respect to the harvest date for a set of crops including sample crops and other crops.

When the training data 13 includes data of sample crops with different reference dates such as fruit bearing dates, for example, the reference dates are included in the training data 13, and a plurality of probability distributions output by the prediction model 16 are used as reference dates. It suffices to shift and synthesize them according to . When the prediction model 16 outputs the probability distribution of the number of weeks, the prediction distribution 17 may indicate the distribution of the number of harvests for the week to which the harvest date belongs. The processing unit 12 may also convert the predicted distribution of the number of harvests of the sample crops into the predicted distribution 17 of the number of harvests of the entire crop set using the sample ratio. For example, the prediction distribution 17 shows that, of 12,000 crops, including sample crops and other crops, 3,600 will be harvested one week, 6,000 will be harvested the next week, and 6,000 will be harvested the next week. Estimated harvest of 2,400 per week.

Then, the processing unit 12 evaluates the degree of similarity between the predicted distribution 17 and the actual distribution indicated by the total number data 14, and determines the timing of stopping the learning process 15 based on the degree of similarity. At the beginning of the learning process 15 , the output of the prediction model 16 approaches the true probability distribution each time the prediction model 16 is updated, and as a result the prediction distribution 17 approaches the total data 14 . On the other hand, when overfitting occurs, the output of the prediction model 16 has an excessively small variance every time the prediction model 16 is updated, moving away from the true probability distribution.

Therefore, for example, the processing unit 12 evaluates the similarity each time the prediction model 16 is updated, detects a peak of the similarity, and stops the learning process 15 when the peak is detected. The prediction model 16 is output as a learning result. The processing unit 12 calculates the error (total error) between the predicted distribution 17 and the actual distribution indicated by the total data 14 as an index indicating the degree of similarity between the two, and detects the timing when the total error becomes the minimum. good too. The error may be the residual sum of squares obtained by summing the squares of the difference between the predicted harvest number and the actual harvest number for each harvest date. Further, the processing unit 12 may stop the learning process 15 when the similarity evaluation result indicates that the predicted distribution 17 and the actual distribution indicated by the total data 14 are similar to each other by a predetermined criterion or more.

According to the machine learning device 10 of the first embodiment, the prediction model 16 that predicts the required number of days from the reference date such as the date of fruiting to the harvest date is generated from the information of the growing environment such as the temperature and the amount of solar radiation. be. Therefore, it is possible to predict the harvest date and the number of crops harvested. Also, the prediction model 16 is learned so as to output the probability distribution of the number of required days instead of the expected value of the number of required days. Therefore, it is possible to express variations in harvest dates, taking into consideration the individual differences in crops that grow at different growth rates even under the same growing environment.

In addition to evaluating the error of the prediction result of the prediction model 16 for each record included in the training data 13, the prediction distribution 17 of the number of harvests predicted from the entire training data 13 and the total number data 14 are shown The degree of similarity between the actual distribution of harvest numbers is evaluated. Based on this degree of similarity, the iteration of updating the prediction model 16 by the learning process 15 is stopped. Therefore, it is possible to prevent the prediction model 16 from outputting a probability distribution with an excessively small variance due to over-learning, and the prediction accuracy of the prediction model 16 is improved.

In particular, collecting detailed information such as reference dates and harvest dates for individual sample crops imposes a heavy burden on farmers. Also, due to individual differences in growth, the number of required days indicated by the training data 13 is biased. If the prediction model 16 is generated using such training data 13, there is a high possibility that the prediction model 16 will output a probability distribution with inappropriate variance when overfitting occurs. On the other hand, according to the machine learning device 10, over-learning can be suppressed, the prediction model 16 can output a probability distribution with an appropriate variance, and it is possible to express variations in harvest dates.

In addition, there is a cross-validation method as a machine learning technique for generating a prediction model with the highest possible prediction accuracy from a small amount of training data. The cross-validation method divides the dataset into multiple blocks (e.g., 10 blocks), selects one of the multiple blocks as test data, and trains the remaining blocks (e.g., 9 blocks). Select as data. A prediction model is generated using training data, and the prediction accuracy of the prediction model is measured using test data. Generation of the prediction model is repeated multiple times (for example, 10 times) by changing the blocks selected as test data.

That is, the cross-validation method finds a combination of records that generates a prediction model with the highest possible prediction accuracy by repeating prediction model generation while replacing records included in training data. However, when the number of available records is very small, it is difficult to properly divide the data set into multiple blocks, and it is not easy to improve the accuracy of crop yield prediction even by the cross-validation method.

[Second embodiment]
Next, a second embodiment will be described.
FIG. 2 is a diagram illustrating an example of an information processing system according to the second embodiment.

The information processing system of the second embodiment uses machine learning to predict the harvest date and the number of crops. Harvest date and harvest number forecasts can be used by farmers as a basis for contracts with shipping destinations. The information processing system of the second embodiment is suitable for managing crops with large individual differences in growth and large variations in harvest dates. In the second embodiment, paprika is assumed as the type of crop. However, the information processing system of the second embodiment can also be applied to management of crops other than paprika.

The information processing system of the second embodiment includes a greenhouse 20, a network 30, a weather data server 31 and a machine learning device 100. FIG.
The interior of the greenhouse 20 includes a sample cultivation area 21 and a general cultivation area 22 as farmland for cultivating paprika. The paprika cultivated in the sample cultivation area 21 is a sample fruit for which a farmer individually observes the date of fruiting and the date of harvest. The paprika cultivated in the general cultivation area 22 is a fruit whose fruiting date and harvest date are not individually observed. The sample fruit cultivated in the sample cultivation area 21 accounts for about 0.1% of the total fruit in the sample cultivation area 21 and the general cultivation area 22 combined. However, for shipping management, the total number of harvests for each harvest day is counted. Moreover, instead of dividing the farmland into the sample cultivation area 21 and the general cultivation area 22, fruits of several trees scattered in the farmland may be selected as sample fruits. Moreover, although one greenhouse is shown in FIG. 2, the farmland may be divided into a plurality of greenhouses.

A sensor 23 is installed inside the greenhouse 20 . The sensor 23 is a sensor device that measures at least temperature and solar radiation. The temperature and amount of solar radiation measured by the sensor 23 are those inside the greenhouse 20, and are different from the temperature and amount of solar radiation outside. The sensor 23 periodically transmits measured data to a predetermined information processing device.

Network 30 includes a wide area data communication network such as the Internet. A weather data server 31 and a machine learning device 100 are connected to the network 30 . A sensor 23 may also be connected to the network 30 .

The weather data server 31 is a server computer that provides weather forecast data indicating the weather forecast for the current day and later. Weather forecast data are provided by public agencies or private weather companies. Weather data server 31 transmits weather forecast data to machine learning device 100 in response to a request from machine learning device 100 . The weather forecast data includes forecast outdoor temperature and forecast solar radiation from the current day onwards. The forecast air temperature and the forecast solar radiation amount are preferably hourly values. The numbers can be daily forecasts such as next day temperature and radiation at 6:00 am, weekly forecasts such as next week's average temperature and radiation at 6:00 am, and next month's averages at 6:00 am. Monthly forecasts such as temperature and average solar radiation are also acceptable.

The machine learning device 100 is a computer that generates a prediction model by machine learning and uses the prediction model to predict the harvest date and harvest number of paprika. The machine learning device 100 collects sample data indicating the fruit bearing date and the harvest date of each sample fruit in the past year (eg, the previous year). The machine learning device 100 also collects weather data indicating the temperature and the amount of solar radiation measured by the sensor 23 . The machine learning device 100 uses sample data and weather data to generate a prediction model that predicts the number of days required from bearing fruit to harvesting of paprika from the temperature and the amount of insolation during the period from bearing fruit to harvesting.

The machine learning device 100 uses a prediction model to predict the number of days required after the paprika of the current year is observed to bear fruit in the sample cultivation area 21 and before the harvest time. At this time, the machine learning device 100 receives weather forecast data from the weather data server 31 . For the temperature and solar radiation corresponding to the input of the prediction model, those measured by the sensor 23 are used for the period before the prediction date, and the weather forecast data are used for the period after the prediction date. The machine learning device 100 predicts the harvest date and number of paprika from the date and number of fruit bearing of the sample fruit, the required number of days output by the prediction model, and the ratio (0.1%) of the sample fruit.

Note that the machine learning device 100 may be a client computer or a server computer. Further, the machine learning device 100 may be a computer owned by a farmer, or may be a computer owned by an information processing business such as a data center. When the machine learning device 100 is owned by a farmer, for example, the machine learning device 100 receives weather data from the sensor 23 without going through a wide area data communication network, and accepts input of sample data from the farmer's user. When the farmer does not own the machine learning device 100, for example, the machine learning device 100 receives weather data and sample data from a terminal device owned by the farmer via a wide area data communication network. Further, in the second embodiment, the machine learning device 100 performs both prediction model generation and harvest prediction using the prediction model, but different computers may perform both. For example, a server computer may generate a forecast model, and a client computer may use the forecast model to make harvest forecasts.

FIG. 3 is a diagram illustrating a hardware example of a machine learning device;
Machine learning device 100 has CPU 101 , RAM 102 , HDD 103 , image interface 104 , input interface 105 , medium reader 106 and communication interface 107 . The above units are connected to a bus. A CPU 101 corresponds to the processing unit 12 of the first embodiment. A RAM 102 or HDD 103 corresponds to the storage unit 11 of the first embodiment. The weather data server 31 and the like also have similar hardware.

The CPU 101 is a processor that executes program instructions. The CPU 101 loads at least part of the programs and data stored in the HDD 103 into the RAM 102 and executes the programs. Note that the CPU 101 may include multiple processor cores, and the machine learning device 100 may include multiple processors. A collection of multiple processors is sometimes called a "multiprocessor" or simply a "processor."

The RAM 102 is a volatile semiconductor memory that temporarily stores programs executed by the CPU 101 and data used by the CPU 101 for calculation. Note that the machine learning device 100 may be provided with a type of memory other than the RAM, and may be provided with a plurality of memories.

The HDD 103 is a nonvolatile storage that stores an OS (Operating System), software programs such as middleware and application software, and data. Note that the machine learning device 100 may include other types of storage such as flash memory and SSD (Solid State Drive), or may include multiple storages.

The image interface 104 outputs an image to the display device 111 connected to the machine learning device 100 according to instructions from the CPU 101 . As the display device 111, any type of display device can be used, such as a CRT (Cathode Ray Tube) display, a liquid crystal display (LCD: Liquid Crystal Display), an organic EL (OEL: Organic Electro-Luminescence) display, or a projector. . Also, an output device other than the display device 111, such as a printer, may be connected to the machine learning device 100. FIG.

Input interface 105 receives an input signal from input device 112 connected to machine learning apparatus 100 . Input device 112 can be any type of input device such as a mouse, touch panel, touch pad, keyboard, or the like. Also, multiple types of input devices may be connected to the machine learning device 100 .

The medium reader 106 is a reading device that reads programs and data recorded on the recording medium 113 . Any type of recording medium can be used as the recording medium 113, such as magnetic disks such as flexible disks (FDs) and HDDs, optical disks such as CDs (Compact Discs) and DVDs (Digital Versatile Discs), and semiconductor memories. can be done. The medium reader 106 copies, for example, programs and data read from the recording medium 113 to other recording media such as the RAM 102 and the HDD 103 . The read program is executed by the CPU 101, for example. Note that the recording medium 113 may be a portable recording medium, and may be used for distribution of programs and data. Also, the recording medium 113 and the HDD 103 may be referred to as a computer-readable recording medium.

The communication interface 107 is connected to the network 30 and communicates with other information processing devices such as the weather data server 31 . The communication interface 107 may be a wired communication interface connected to a wired communication device such as a switch or router, or a wireless communication interface connected to a wireless communication device such as a base station or access point.

Next, a method of harvest prediction using a prediction model will be described. Note that in the second embodiment, observation of fruiting and management of harvest are performed on a weekly basis. Therefore, the date of bearing fruit and the date of harvest of the sample data are the dates of a specific day of the week. Moreover, the required number of days output by the prediction model indicates the number of weeks, and the predicted harvest date is the date of a specific day of the week.

FIG. 4 is a diagram showing an example of a data flow for harvest prediction.
When one or more sample fruits are observed to bear on a certain day, a date of bearing of fruits 211 and a number of specimens of fruit bearing 212 are collected as sample data. The date of bearing fruit 211 is the date on which the farmer observed the bearing of fruits, and the number of specimen bearing fruits 212 is the number of specimen bearing fruits on that day. For example, the fruit bearing date 211 is October 23, and the sample number of fruit bearing 212 is five.

Then, the average temperature 213 and the average amount of solar radiation 214 indoors of the greenhouse 20 are calculated as explanatory variables for the period from the fruit bearing date 211 to the harvest time. The average temperature 213 and the average amount of solar radiation 214 are values for each hour. Therefore, explanatory variables are 48-dimensional vectors. In the second embodiment, the average temperature and the average amount of solar radiation are used as explanatory variables, but other indicators such as the cumulative temperature and the cumulative amount of solar radiation may be used.

The average temperature 213 and the average amount of solar radiation 214 are calculated as follows. For the period from the date of bearing fruit to the day before the predicted date, the measured temperature 221 and the amount of solar radiation 222 inside the greenhouse 20 measured by the sensor 23 are used. For the period from the forecast date to the harvest time, forecast temperature 223 and forecast solar radiation 224 of the weather forecast data are used.

However, the forecast air temperature 223 and the forecast solar radiation amount 224 are outdoor air temperature and solar radiation amount. Therefore, using the environmental parameter 227 , the predicted temperature 223 is converted into the predicted indoor temperature 225 of the greenhouse 20 , and the predicted solar radiation amount 224 is converted into the predicted indoor solar radiation amount 226 of the greenhouse 20 . The environmental parameters 227 indicate the relationship between outdoor temperature and indoor temperature, and the relationship between outdoor solar radiation and indoor solar radiation. For example, environmental parameters 227 include a linear expression for converting outdoor temperature to indoor temperature and a linear expression for converting outdoor solar radiation to indoor solar radiation. Environmental parameters 227 are prepared in advance. The environmental parameters 227 may be individual parameters individually adjusted for each greenhouse, or may be general parameters commonly applied to various greenhouses.

Measured temperature 221 and expected temperature 225 are averaged to calculate average temperature 213 , and measured solar radiation 222 and expected solar radiation 226 are averaged to calculate average solar radiation 214 . Then, an average temperature 213 and an average amount of solar radiation 214 are input to a prediction model 210 generated in advance, and a required number of days 215 is output from the prediction model 210 . The required number of days 215 is an estimate of the number of days from fruit setting to harvest. The harvest date 216 is calculated by adding the required number of days 215 to the fruit bearing date 211 . The harvest date 216 is a prediction of a suitable harvest date for the fruit that bears fruit on the fruit bearing date 211 . For example, the number of days required 215 is eight weeks and the harvest date 216 is December 18th.

Also, the sample harvest number 217 is calculated from the sample fruit bearing number 212 . The number of harvested samples 217 is a prediction of the number of sample fruits that will be harvested on the harvest date 216 among the sample fruits observed on the fruit bearing date 211 . Here, it is assumed that the prediction model 210 outputs the expected value of the required number of days as the required number of days 215 , so the number of sample harvests 217 is the same as the number of sample fruit bearings 212 . For example, the sample harvest number 217 is five. However, as will be described later, it is also possible to generate a prediction model that outputs a probability distribution of required days. In that case, the sample harvest number 217 indicates the number of sample fruits for each required number of days. The number of sample fruits for each required number of days can be calculated by multiplying the sample fruit bearing number 212 by the probability for each required number of days.

Then, the sample harvest number 217 is converted to the harvest number 218 . The number of harvests 218 is a prediction of the number of fruits that are predicted to bear on the date 211 of fruiting and will be harvested on the date 216 of harvesting. The fruit predicted to bear fruit on the fruit-bearing date 211 includes observed sample fruit and other fruits. The number of harvests 218 is calculated from the number of sample harvests 217 and the sample ratio 219 . The sample ratio 219 is the ratio of the sample fruit to the whole fruit. The harvested number 218 can be calculated by dividing the sample harvested number 217 by the sample ratio 219 , that is, by multiplying the sample harvested number 217 by the reciprocal of the sample ratio 219 . For example, the sample rate 219 is 0.1%, and the number of harvests 218 is 500/0.1%=500×1,000=5,000.

In this way, it is predicted that 218 fruits will be harvested on the harvest date 216 for the fruits that bear fruit on the fruit bearing date 211 . For example, for a fruit set on October 23rd, 5,000 fruits are expected to be harvested on December 18th. By summing the predictions for different fruiting dates, it is possible to predict the overall harvest date and number of crops.

However, paprika has a large individual difference in growth, so even if the fruiting date is the same, the harvest date varies. Therefore, if the prediction model 210 that outputs the expected value of the required number of days is used, there is a risk that the prediction of the overall harvest date and number of harvests will deviate from the actual situation.

FIG. 5 is a diagram showing a usage example of a prediction model that outputs expected values.
Suppose that sample fruit bearing numbers 231, 232, and 233 are measured on different fruit bearing days. The number of sample fruit bearing 231 indicates 5 sample fruits for which fruit bearing was observed on October 23rd. The number of sample fruit bearing 232 indicates three sample fruit for which fruit bearing was observed on October 30th. The sample fruit bearing number 233 indicates 4 sample fruits on which fruit bearing was observed on November 6th.

The required number of days is predicted for each of the sample fruit bearing numbers 231 , 232 , and 233 . Here, it is assumed that a prediction model that outputs the expected value of required days is used. Then, the expected value of the required number of days is calculated from the average temperature and the average amount of solar radiation after October 23 for the number of sample fruits set 231 . For the number of sample fruit bearings 232, the expected value of the required number of days is calculated from the average temperature and the average amount of solar radiation after October 30th. For the sample number of fruit bearing 233, the expected value of the required number of days is calculated from the average temperature and the average amount of solar radiation after November 6th. Since different average temperatures and average insolation are used for different fruiting days, different duration expectations can be calculated. Here, the required number of days for the sample number 231 of bearing fruit is 8 weeks, the required number of days for the sample number 232 of bearing fruit is 7 weeks, and the required number of days for the sample number 233 of bearing fruit is 6 weeks.

Then, sample harvest numbers 234, 235, and 236 are predicted. The sample harvest number 234 indicates five sample fruits expected to be harvested on December 18, eight weeks after October 23rd. Sample Harvest Number 235 indicates three sample fruits expected to be harvested on December 18, seven weeks after October 30th. The sample harvest number 236 indicates four sample fruits expected to be harvested on December 18, six weeks after November 6th. Summing up the sample harvest numbers of 234, 235, and 236, 0 sample fruits were harvested on December 11, 12 sample fruits were harvested on December 18, and 0 sample fruits were harvested on December 25. It is predicted that it will be harvested.

Converting these sample yields to total yields using a sample fraction of 0.1% predicts yields of 237, 238, and 239. Harvest number 237 indicates 0 fruits expected to be harvested on December 11th. Harvest number 238 indicates 12,000 fruits expected to be harvested on December 18th. Harvest number 239 indicates 0 fruits expected to be harvested on December 25th. In this way, using a prediction model that outputs the expected value of the required number of days may result in a prediction that the number of harvests will be concentrated on a specific harvest date. However, the harvest numbers 237, 238, and 239 are less reliable because the harvest dates actually vary due to individual differences. Therefore, we use a prediction model that outputs the probability distribution of the number of required days instead of the expected number of required days.

FIG. 6 is a diagram showing a usage example of a prediction model that outputs a probability distribution.
Probability distributions of the number of required days are predicted for each of the number of sample fruit bearings 231, 232, and 233. FIG. A probability distribution of the required number of days is calculated from the average temperature and the average amount of solar radiation after October 23 for the sample fruit bearing number 231 . A probability distribution of the required number of days is calculated from the average temperature and the average amount of solar radiation after October 30 for the sample number of fruits 232 . A probability distribution of the required number of days is calculated from the average temperature and the average amount of solar radiation after November 6th for the sample fruit bearing number 233 .

Since different average temperatures and average insolation are used for different fruiting days, probability distributions for different required days can be calculated. Here, the probability distribution with respect to the sample number of bearing fruit 231 is 30% for 7 weeks, 50% for 8 weeks, and 20% for 9 weeks. The probability distribution for the sample number of fruit bearing 232 is 30% for 6 weeks, 50% for 7 weeks, and 20% for 8 weeks. The probability distribution for the number of specimen bearing 233 is 30% for 5 weeks, 50% for 6 weeks, and 20% for 7 weeks.

Then, a distribution of 241, 242, and 243 sample harvests is predicted for the sample fruit bearing number 231 on October 23rd. Sample harvest number 241 indicates 5 x 30% = 1.5 sample fruits expected to be harvested on December 11th. The number of sample harvests 242 indicates 5 x 50% = 1.5 sample fruits expected to be harvested on December 18th. Sample harvest number 243 indicates 5×20%=1.0 sample fruits expected to be harvested on December 25th.

Similarly, a distribution of 244, 245, and 246 harvested specimens is predicted for 232 specimens set on October 30th. The sample harvest number 244 indicates 3 x 30% = 0.9 sample fruit expected to be harvested on December 11th. Sample Harvest 245 indicates 3 x 50% = 1.5 sample fruits expected to be harvested on December 18th. The number of sample harvests 246 indicates 3 x 20% = 0.6 sample fruits expected to be harvested on December 25th.

Also, a distribution of 247, 248, and 249 sample harvests is predicted for 233 sample fruit bearings on November 6th. Sample harvest number 247 indicates 4 x 30% = 1.2 sample fruits expected to be harvested on December 11th. Sample Harvest 248 indicates 4 x 50% = 2.0 sample fruits expected to be harvested on December 18th. Sample Harvest 249 indicates 4 x 20% = 0.8 sample fruit expected to be harvested on December 25th.

Summing up the sample harvest numbers of 241, 244, and 247 yields a prediction that 3.6 sample fruit will be harvested on December 11th. Summing the sample harvest numbers of 242, 246, and 248 yields a prediction that 6.0 sample fruit will be harvested on December 18th. Summing up the numbers of sample harvests of 243, 246, and 249, it is predicted that 2.4 sample fruits will be harvested on December 25th. Converting these sample yields to total yields using a sample proportion of 0.1% predicts yields of 251, 252, and 253. Harvest number 251 indicates 3,600 fruits expected to be harvested on December 11th. Harvest number 252 indicates 6,000 fruits expected to be harvested on December 18th. Harvest number 253 indicates 2,400 fruits expected to be harvested on December 25th. In this way, by using a prediction model that outputs a probability distribution, variations in harvest dates can be expressed, and the reliability of the harvest numbers 251, 252, and 253 is increased.

Here, the problem is how to learn the prediction model that outputs the probability distribution of the required number of days. Typical machine learning repeatedly evaluates the error in the output of a prediction model using training data and updates the coefficients of the prediction model so as to reduce the error. Various machine learning models such as GP models, multiple regression models, and neural networks can be used as prediction models. While the number of iterations is small, the output error of the prediction model is large and the fitting accuracy to the training data is low. As the number of iterations increases, the output error of the prediction model decreases and the fitting accuracy to the training data increases. The above is often repeated until the error on the training data is small enough.

On the other hand, in the case of crop yield prediction, observation and tracking of individual fruit samples is a heavy burden on farmers, so the number of sample fruit is limited and the amount of training data that can be used for machine learning is small. In addition, since paprika has a large individual difference in growth, the number of days required for these small sample fruits does not accurately represent the variation in the number of days required for all harvested fruits. For this reason, if the number of iterations is increased until the error with respect to the training data becomes sufficiently small, overfitting, in which the prediction model excessively fits the training data, tends to occur. A predictive model that is overtrained from a small amount of training data will output a probability distribution with excessively small variance. As a result, the reliability of the probability distribution output by the prediction model is lowered.

FIG. 7 is a diagram illustrating a usage example of an under-learning prediction model.
Consider an early-stage predictive model with a small number of iterations. The probability distribution output by the insufficiently trained prediction model has a large variance because the number of required days is not sufficiently narrowed down.

For October 23rd fruit set, the prediction model outputs a probability distribution of 33% for 7 weeks, 33% for 8 weeks, and 33% for 9 weeks. Then, since the number of specimen bearing fruit 231 is 5, the number of harvested specimens is predicted to be 1.7 on December 11th, 1.7 on December 18th, and 1.7 on December 25th. be. Similarly, for October 30th fruit set, the prediction model outputs a probability distribution of 33% for 6 weeks, 33% for 7 weeks, and 33% for 8 weeks. Then, since the number of specimen bearing fruit 232 is 3, the number of harvested specimens is predicted to be 1.0 on December 11th, 1.0 on December 18th, and 1.0 on December 25th. be. For November 6th fruit set, the prediction model outputs a probability distribution of 33% for 5 weeks, 33% for 6 weeks, and 33% for 7 weeks. Then, since the number of specimen bearing fruit 233 is 4, the number of harvested specimens is predicted to be 1.3 on December 11th, 1.3 on December 18th, and 1.3 on December 25th. be.

Summing up the numbers of harvested specimens for each harvest date yields 4.0 on December 11th, 4.0 on December 18th, and 4.0 on December 25th. Using a sample proportion = 0.1%, we predict an overall yield of 254,255,256. Harvest number 254 indicates 4,000 as the harvest number on December 11th. Harvest number 255 indicates 4,000 as the harvest number on December 18th. Harvest number 255 indicates 4,000 as the harvest number on December 25th.

In this way, if an under-trained prediction model is used, the variance of the probability distribution becomes excessively large, and the required number of days is not properly narrowed down. As a result, the predicted harvest numbers 254, 255, and 256 fluctuate excessively and reliability decreases.

FIG. 8 is a diagram showing an example of using an over-trained prediction model.
Consider an overtrained prediction model with many iterations. The probability distribution output by the over-learned prediction model is too well adapted to the required number of days indicated by the training data and has a small variance.

For October 23rd fruit set, the prediction model outputs a probability distribution of 0% for 7 weeks, 100% for 8 weeks, and 0% for 9 weeks. Then, since the number of specimen bearing fruit 231 is 5, the number of harvested specimens is predicted to be 0 on December 11th, 5 on December 18th, and 0 on December 25th. For October 30th fruit set, the prediction model outputs a probability distribution of 0% for 6 weeks, 100% for 7 weeks, and 0% for 8 weeks. Then, since the number of specimen bearing fruit 232 is 3, the number of harvested specimens is predicted to be 0 on December 11th, 3 on December 18th, and 0 on December 25th. For fruit set on Nov. 6, the prediction model outputs a probability distribution of 0% for 5 weeks, 100% for 6 weeks, and 0% for 7 weeks. Then, since the number of specimen bearing fruit 233 is 4, the number of harvested specimens is predicted to be 0 on December 11th, 4 on December 18th, and 0 on December 25th.

Summing up the numbers of harvested specimens for each harvest date yields 0 specimens on December 11th, 12 specimens on December 18th, and 0 specimens on December 25th. Using a sample proportion = 0.1%, we predict an overall yield of 257,258,259. The number of harvests 257 indicates 0 as the number of harvests on December 11th. Harvest number 258 indicates 12,000 as the harvest number on December 18th. The number of harvests 259 indicates 0 as the number of harvests on December 25th.

In the above example, the harvest numbers 257, 258 and 259 are the same as the harvest numbers 237, 238 and 239 shown in FIG. In other words, even if a prediction model that outputs a probability distribution is used, if the variance becomes excessively small due to overfitting, prediction results close to those of the prediction model that outputs the expected value will be obtained. reliability does not improve.

The variance of the probability distribution output by the prediction model decreases as the number of machine learning iterations increases. Therefore, by stopping machine learning iterations at an appropriate number of times, it is possible to guide the variance of the probability distribution to an appropriate size. Therefore, the question arises as to when to stop machine learning iterations.

Here, regarding paprika cultivation in past years, sample data indicating the actual number of days required from fruit setting to harvest is collected only for a small number of sample fruits, while total data indicating the total number of harvests for each harvest date is collected. are collected using agricultural machinery for shipping management. Therefore, the machine learning device 100 predicts the total number of harvests in the past year from the prediction model at that time, the training data, and the sample ratio for each iteration that updates the coefficients of the prediction model, and compares the prediction and the actual results indicated by the total data. The comparison determines when to stop the iteration. The total yield can be predicted by applying a method similar to that of FIG. 6 to the training data.

If the variance of the probability distribution output by the prediction model is excessively large, the prediction of the total number of harvests is likely to be dissimilar to the actual results. Also, when the variance of the probability distribution output by the prediction model is excessively small, there is a high possibility that the prediction of the total number of harvests will not resemble the actual results. On the other hand, when the variance of the probability distribution output by the prediction model is optimal reflecting the variation in the actual harvest dates, it is likely that the degree of similarity between the prediction of the total number of harvests and the actual number of harvests will be maximized. Therefore, the machine learning device 100 adopts the prediction model with the maximum similarity as the learning result.

FIG. 9 is a diagram illustrating an example of stop timing of machine learning.
As the number of iterations increases, the probability distributions output by the prediction model for a specific average temperature and average solar radiation amount change as shown in probability distributions 261, 262, and 263. FIG.

A probability distribution 261 is output from the under-learning prediction model, and corresponds to the prediction model in FIG. That is, the variance of probability distribution 261 is excessively large. Probability distribution 262 is output from the optimal prediction model and corresponds to the prediction model of FIG. That is, the variance of the probability distribution 262 is optimal reflecting the variation in paprika harvest dates. Probability distribution 263 is output from the overtrained prediction model and corresponds to the prediction model of FIG. That is, the variance of probability distribution 263 is too small.

When the prediction model outputs the probability distribution 261, the machine learning device 100 predicts the harvest number distribution 264 from the training data in the same manner as in FIG. The harvest number distribution 264 shows the forecast of the total number of harvests by harvest date. The harvest number distribution 264 corresponds to the harvest numbers 254, 255, and 256. That is, the machine learning device 100 calculates the probability distribution of the number of harvests by inputting the average temperature and the average amount of solar radiation into the prediction model for each record of the training data, multiplies the probability distribution by the sample number of fruit bearing, and Calculate the number of harvested specimens. The machine learning device 100 totals the predictions for each record of the training data and multiplies them by the reciprocal of the sample ratio to calculate the number of harvests for each harvest date.

When the harvest number distribution 264 is predicted, the machine learning device 100 compares the harvest number distribution 264 and the harvest number distribution 267 to calculate an error (total error). The harvest number distribution 267 is the harvest situation in the same year as the training data, and indicates the overall harvest number for each harvest date. Harvest number distribution 267 indicates that the number of harvests on December 11 is 3,700, the number of harvests on December 18 is 5,800, and the number of harvests on December 25 is 2,500. As an index of total error, for example, residual sum of squares is used. The residual sum of squares is a numerical value obtained by calculating the square of the difference in the number of harvests between the forecast and the actual number of harvests for each harvest date and summing the squares of the difference. The sum of squared residuals of the harvest number distribution 264 and the harvest number distribution 267 is 5,580,000. Therefore, the total error is large.

Next, when the prediction model outputs the probability distribution 262, the machine learning device 100 predicts the harvest number distribution 265 from the training data in the same manner as in FIG. The harvest number distribution 265 corresponds to the harvest numbers 251, 252, and 253. When the harvest number distribution 265 is predicted, the machine learning device 100 compares the harvest number distribution 265 and the harvest number distribution 267 to calculate the total error. The sum of squared residuals of the harvest number distribution 265 and the harvest number distribution 267 is 60,000. Therefore, the total error is smaller than when the prediction model outputs probability distribution 261 .

Next, when the prediction model outputs the probability distribution 263, the machine learning device 100 predicts the harvest number distribution 266 from the training data in the same manner as in FIG. The harvest number distribution 266 corresponds to harvest numbers 257, 258, and 259. When the harvest number distribution 266 is predicted, the machine learning device 100 compares the harvest number distribution 266 and the harvest number distribution 267 to calculate the total error. The sum of squared residuals of the harvest number distribution 266 and the harvest number distribution 267 is 58,380,000. Therefore, the total error is greater than when the prediction model outputs probability distribution 262 .

In this way, the machine learning device 100 detects that the total error is minimized when the prediction model outputs the probability distribution 262, that is, the similarity is maximized. Then, the machine learning device 100 stops the machine learning iterations and outputs a prediction model that outputs the probability distribution 262 as a learning result.

FIG. 10 is a diagram illustrating an example of a machine learning data flow.
The training data used to generate the predictive model 270 includes multiple records with different fruit bearing dates. Each record of the training data includes a fruit-bearing date 271 , sample fruit-bearing number 272 , sample number of days distribution 273 , average temperature 277 and average solar radiation 278 . The number of sample days distribution 273 indicates the number of sample harvests for each required number of days. The sample number of days distribution 273 may be represented by the number, or may be represented by the probability obtained by dividing the number by the sample number of fruits 272 . For example, the sample days distribution 273 indicates that 7 weeks is 40%, 8 weeks is 60%, and 9 weeks is 0%.

The average temperature 277 is the hourly indoor temperature, which is averaged over the period from the fruit bearing date 271 to the harvest date. The average solar radiation amount 278 is the hourly indoor solar radiation amount, which is averaged over the period from the fruit bearing date 271 to the harvest date. Therefore, the average temperature 277 and the average amount of solar radiation 278 are each 24-dimensional vectors, making a total of 48-dimensional vectors. Average air temperature 277 is calculated from measured air temperature 275 measured by sensor 23 . Average solar radiation 278 is calculated from measured solar radiation 276 measured by sensor 23 . Since the training data shows sample fruits from previous years, the measured temperature 275 and the measured solar radiation amount 276 from the date of fruiting 271 to the date of harvest are already known, and weather forecast data need not be used.

In addition to the training data, total data indicating the number of harvests 274 is prepared in advance. The number of harvests 274 is the actual number of harvests for each harvest date. For example, a harvest number of 274 means that 3,700 fruits were harvested on December 11th, 5,800 fruits were harvested on December 18th, and 2,500 fruits were harvested on December 25th. indicates that the

When machine learning starts, the coefficients of predictive model 270 are initialized. For each record of training data, the prediction model 270 receives an average temperature 277 and an average amount of solar radiation 278, and outputs a required number of days distribution 281 from the prediction model 270. FIG. The required number of days distribution 281 indicates the prediction of the harvest probability for each required number of days. For example, the required days distribution 281 indicates that 7 weeks is 33%, 8 weeks is 33%, and 9 weeks is 33%. For each record of training data, the required number of days distribution 281 and the sample number of days distribution 273 are compared to calculate an error. Then, the errors for each record of the training data are summed to calculate the model error 282 for the entire training data.

For the error for each record of training data, for example, residual sum of squares is used. This residual sum of squares is an index obtained by squaring the difference between the value of the distribution of required days 281 and the value of the sample distribution of days 273 for each required number of days, and totaling the squares of the differences for a plurality of required days. The comparison of the required number of days distribution 281 and the sample number of days distribution 273 may be performed as a comparison of probabilities such as a comparison of 33% and 40%. Further, the comparison between the required number of days distribution 281 and the sample number of days distribution 273 may be performed by multiplying the probability by the sample number of fruits 272 to compare numbers such as 1.7 and 2.

After the model error 282 is calculated, the coefficients of the prediction model 270 are updated so that the model error 282 becomes smaller. When updating the coefficients of the prediction model 270, the previous coefficient is saved. The process from updating the prediction model 270 to calculating the model error 282 is one iteration. Updating of the prediction model 270 is repeated until it is decided to stop the iteration by the following stop determination. Stop judgment is executed for each iteration. The stop determination may be performed by interrupting the iteration after the forecast model 270 outputs the required days distribution 281 until the forecast model 270 is updated next time. Also, the stop determination may be executed in parallel with the above iterations. Different processors or processor cores may be used to perform iteration and stop determination in parallel.

After the required number of days distribution 281 is calculated, a sample harvest number 283 is calculated by multiplying the harvest probability by the sample fruit bearing number 272 for each record of the training data. The number of sample harvests 283 indicates the prediction of the number of sample fruit harvests for each required number of days. For example, the number of harvested specimens of 283 indicates that 1.7 fruits were harvested in 7 weeks, 1.7 in 8 weeks, and 1.7 in 9 weeks out of 5 specimen fruits.

The number of days required for the sample harvest number 283 for each record of the training data is shifted based on the fruit bearing date 271 so that the harvest dates are aligned. For example, 7 weeks after October 23rd corresponds to 6 weeks after October 30th. The sample harvest number 283 corresponding to the record that is the 23rd of the month is shifted one week later. For a plurality of records of training data, sample harvest numbers 283 with the same harvest date are totaled for each harvest date.

Then, the harvested number 285 is calculated by multiplying the reciprocal of the sample rate 284 by the summed number of harvested samples (divided by the sample rate 284). For example, the combined number of sample harvests is multiplied by 1,000. Harvest number 285 indicates the forecast of the number of whole fruit harvests for each harvest date. For example, a harvest number of 285 means that 4,000 fruits were harvested on December 11th, 4,000 fruits were harvested on December 18th, and 4,000 fruits were harvested on December 25th. It shows the prediction that it will be done.

After the harvested number 285 is calculated, the harvested number 285 and the harvested number 274 are compared to calculate the total error 286 . For the total error 286, for example, residual sum of squares is used. Then, the total error 286 of the previous iteration and the total error 286 of the current iteration are compared. If the current total error 286 is less than or equal to the previous total error 286, it is decided to continue the iteration. In this case, predictive model 270 is updated according to model error 282 .

On the other hand, if the current total error 286 is greater than the previous total error 286, it is decided to stop the iteration. In this case, predictive model 270 is not updated. Since the coefficients of the optimum prediction model 270 are the coefficients of the previous iteration, the saved coefficients of the prediction model 270 are read out and output as the learning result. That is, it is detected that the similarity between the number of harvests 274 and 285 is maximized and the total error 286 is minimized. Here, it is assumed that the total error 286 monotonically decreases before reaching the optimal prediction model 270 coefficients, and that the total error 286 monotonically increases after reaching the optimal prediction model 270 coefficients.

Next, functions of the machine learning device 100 will be described.
FIG. 11 is a block diagram illustrating an example of functions of the machine learning device.
Machine learning device 100 includes weather data storage unit 121, sample data storage unit 122, total data storage unit 123, prediction model storage unit 124, data collection unit 125, data processing unit 126, machine learning unit 127, iteration control unit 128, and It has a harvest prediction unit 129 . The weather data storage unit 121, the sample data storage unit 122, the total data storage unit 123, and the prediction model storage unit 124 are realized using storage areas of the RAM 102 or the HDD 103, for example. The data collection unit 125, the data processing unit 126, the machine learning unit 127, the iteration control unit 128, and the harvest prediction unit 129 are realized using programs, for example.

The meteorological data storage unit 121 stores meteorological data from the date of bearing fruit in the past year to the date of harvest, and meteorological data from the date of bearing fruit in the current year to the day before the forecast date. The meteorological data includes measured air temperature and measured solar radiation measured by sensor 23 . In addition, the weather data storage unit 121 stores weather forecast data after the forecast date of the current year. Weather forecast data is collected from the weather data server 31 . Weather forecast data includes outdoor forecast temperatures and forecast solar radiation. The weather data storage unit 121 also stores environmental parameters for converting outdoor forecast air temperature and forecast solar radiation into indoor forecast air temperature and forecast solar radiation.

The sample data storage unit 122 stores sample data indicating the fruit bearing date and harvest date for each sample fruit in the past year and sample data indicating the fruit bearing date for each sample fruit in the current year. The sample data storage unit 122 also stores a sample ratio, which is the ratio of the sample fruit to the whole fruit.

The total number data storage unit 123 stores total number data indicating the number of harvests for each harvest date in the past year.
The prediction model storage unit 124 stores prediction models as learning results.
The data collection unit 125 collects various data stored in the weather data storage unit 121 , sample data storage unit 122 and total number data storage unit 123 . As a data collection method, the data collection unit 125 may receive data input from the user. The data collection unit 125 may also receive data from other information processing devices.

The data processing unit 126 processes past year weather data stored in the weather data storage unit 121 and past year sample data stored in the sample data storage unit 122, and performs training including a plurality of records with different fruit bearing dates. Generate data. Specifically, the data processing unit 126 extracts the date of bearing fruit from the sample data of the previous year, counts the number of sample bearing days for each date of bearing fruit, and calculates the sample number of days from the difference between the date of bearing fruit and the date of harvest for each bearing date. Calculate the distribution. The data processing unit 126 also extracts the measured temperature and amount of solar radiation from the date of fruiting to the date of harvest from the meteorological data of past years, and calculates the average temperature and amount of solar radiation for each hour.

In addition, the data processing unit 126 processes the current year's weather data and weather forecast data stored in the weather data storage unit 121 and the current year's sample data stored in the sample data storage unit 122 to produce data for harvest prediction. generate the input data for Specifically, the data processing unit 126 extracts the date of fruit bearing from the sample data of the current year, and counts the number of sample fruit bearings for each date of fruit bearing. In addition, the data processing unit 126 extracts the measured temperature and the measured amount of solar radiation from the date of fruiting to the day before the predicted date for each day of fruiting from the weather data of the current year. The data processing unit 126 extracts the predicted temperature and solar radiation amount from the forecast date to the harvest time from the weather forecast data, and uses the environmental parameters stored in the weather data storage unit 121 to calculate the predicted indoor temperature and solar radiation. Convert to quantity. Then, the data processing unit 126 calculates the average temperature and the average amount of solar radiation for each hour from the day of fruiting to the harvest time for each day of fruiting.

The data processing unit 126 provides training data to the machine learning unit 127 . The data processing unit 126 also provides the total number data stored in the total number data storage unit 123 to the iteration control unit 128 . The data processing unit 126 provides input data to the harvest prediction unit 129 .

The machine learning unit 127 performs machine learning using training data including multiple records for different fruit bearing dates. The machine learning algorithm to use is specified in advance. The generated prediction model outputs a probability distribution of the required number of days from fruit setting to harvest. The machine learning unit 127 repeats updating the coefficients of the prediction model and calculating the model error for the training data. When the iteration control unit 128 instructs the machine learning unit 127 to stop the iteration, the machine learning unit 127 outputs the previous prediction model to the prediction model storage unit 124 .

Each time the machine learning unit 127 updates the prediction model, the iteration control unit 128 predicts the total number of harvests for each harvest date in the past year from the required number of days distribution, the sample number of fruit bearing, and the sample ratio output by the prediction model, Compare with the performance indicated by the total data. The iteration control unit 128 calculates the total error between the predicted and actual harvest numbers, and instructs the machine learning unit 127 to stop the iteration when the total error increases from the previous time.

The harvest prediction unit 129 predicts the number of harvests for each harvest date in the current year based on the prediction model stored in the prediction model storage unit 124 and the input data provided from the data processing unit 126 . Specifically, the harvest prediction unit 129 inputs the average temperature and the average amount of solar radiation for the current year into the prediction model, and predicts the required number of days distribution for each fruit-bearing day. The harvest prediction unit 129 calculates the harvest date by adding the required number of days to the fruit-bearing date, multiplies the probability indicated by the required number of days distribution by the sample number of fruit-bearing, calculates the number of sample harvests, and multiplies the reciprocal of the sample ratio. Convert to harvest number. The harvest prediction unit 129 sums up the number of harvests on different fruiting days for each harvest date, and predicts the total number of harvests for each harvest date.

The harvest prediction unit 129 outputs a prediction result of the total number of harvests for each harvest date. For example, the harvest prediction unit 129 displays the prediction result on the display device 111 . Also, for example, the harvest prediction unit 129 saves the prediction result in a non-volatile storage such as the HDD 103 . Also, for example, the harvest prediction unit 129 outputs the prediction result to another output device such as a printer. Also, for example, the harvest prediction unit 129 transmits prediction results to other information processing devices.

FIG. 12 is a diagram showing a table example of weather data, sample data, and total number data.
The weather data table 131 is stored in the weather data storage unit 121 . The weather data table 131 includes past year weather data. This year's weather data and weather forecast data can also be managed in a table similar to the weather data table 131 . The weather data table 131 includes items of date and time, temperature, and amount of solar radiation. The date and time are in increments of one hour. Temperature is the hourly average of temperature. The temperature unit is, for example, degrees Celsius. Insolation is an hourly average of instantaneous insolation. The unit of solar radiation is, for example, kW/m ² .

A sample data table 132 is stored in the sample data storage unit 122 . The sample data table 132 includes past year sample data. This year's sample data can also be managed in a table similar to the sample data table 132 . However, the harvest date is not registered for this year's sample data. The sample data table 132 includes items of variety, fruit number, fruiting date, and harvest date. The cultivars are paprika cultivars, including red, yellow and orange cultivars with different fruit colors. Yield prediction is made for each variety. The fruit number is an identification number that individually identifies the sample fruit. A unique fruit number is assigned to the sample fruit within the same variety. The date of bearing fruit is the date on which the bearing of the sample fruit was observed. The harvest date is the date on which the sample fruit was harvested. However, for the convenience of data management, the date of fruiting and the date of harvest are dates of specific days of the week.

The total data table 133 is stored in the total data storage unit 123 . The total data table 133 includes past year total data. The total data table 133 includes items of harvest date, red count, yellow count, and orange count. The harvest date is the date on which the fruit was harvested. However, for convenience of data management, the harvest date is the date of a specific day of the week. The red number is the harvested number of red varieties. The number of yellow cultivars is the harvested number of yellow cultivars. The number of oranges is the harvested number of orange varieties.

FIG. 13 is a diagram showing an example of a training data table.
A training data table 134 is generated based on the weather data table 131 and the sample data table 132 and used for machine learning. The training data table 134 includes items of fruit bearing date, sample fruit bearing number, objective variable, and explanatory variable. The date of bearing fruit is the date of bearing fruit that appears in the sample data table 132 . The number of fruit bearing samples is the number of fruit samples registered in the sample data table 132 that have the same fruit bearing date.

The target variable is the sample day distribution. The sample day distribution is the number of sample fruits for each required number of days, such as 0 for 6 weeks, 2 for 7 weeks, and 3 for 8 weeks. The required number of days is the difference between the date of bearing fruit and the date of harvest in the sample data table 132 . The sample fruit for each required number of days is the sample fruit registered in the sample data table 132 that has the same required number of days. The total number of sample fruit set for each required number of days matches the number of sample fruit set.

Explanatory variables include hourly average temperature and hourly average solar radiation. The hourly average temperature is obtained by extracting the temperature of each date from the date of fruiting to the date of harvest from the weather data table 131, classifying them by time such as 0:00, 1:00, 2:00, . . . and averaging them. It is calculated by The average amount of solar radiation for each hour is obtained by extracting the amount of solar radiation on each date from the date of fruiting to the date of harvest from the weather data table 131, classifying it by time such as 0:00, 1:00, 2:00, . . . It is calculated by converting The harvest date corresponding to a certain fruit-bearing date may be the last harvest date, the first harvest date, or the central harvest date among the harvest dates on which one or more sample fruits are harvested in the sample date distribution. good.

Next, a processing procedure of the machine learning device 100 will be described.
FIG. 14 is a flowchart illustrating an example of machine learning procedures.
(S10) The data collection unit 125 collects weather data, sample data, and total number data. Machine learning is performed for each breed. However, the breed may be added to the explanatory variables of the prediction model.

(S11) The data processing unit 126 extracts the fruit-bearing date from the sample data, and classifies the sample fruit by the fruit-bearing date. The data processing unit 126 counts sample fruits for each fruit-bearing date and uses it as the sample fruit-bearing number of the training data. In addition, the data processing unit 126 calculates the number of required days, which is the difference between the date of bearing fruit and the date of harvest, and counts the number of sample fruits for each required number of days to obtain the sample number of days distribution of the training data.

(S12) The data processing unit 126 extracts the temperature and the amount of solar radiation from the date of fruiting to the date of harvest from the weather data for each day of fruiting. The data processing unit 126 classifies the extracted temperatures by time, and sets the average of the temperatures at each time as the average temperature of the training data. The data processing unit 126 also classifies the extracted solar radiation amounts by time, and sets the average of the solar radiation amounts at each time as the average solar radiation amount of the training data.

(S13) The machine learning unit 127 initializes the coefficients of the prediction model.
(S14) The machine learning unit 127 selects one training data record.
(S15) The machine learning unit 127 inputs the data of the 48-dimensional explanatory variables indicating the average temperature and the average amount of solar radiation to the prediction model, and reads out the data of the objective variable from the prediction model to predict the required number of days distribution. The required number of days distribution indicates the probability for each required number of days.

(S16) The machine learning unit 127 determines whether all the training data records have been selected in step S14. If all the records have been selected, the process proceeds to step S17, and if there are unselected records in the training data, the process returns to step S14.

(S17) The machine learning unit 127 compares the required number of days distribution predicted in step S15 with the sample number of days distribution for each record of the training data to calculate an error. The error is, for example, the residual sum of squares. The machine learning unit 127 calculates model errors for the entire training data. For example, the model error is the sum of errors for each record of the training data.

FIG. 15 is a flowchart (continued) showing a procedure example of machine learning.
(S18) The iteration control unit 128 selects one training data record.
(S19) The iteration control unit 128 multiplies the probability of the required number of days distribution calculated by the machine learning unit 127 in step S15 by the sample number of fruits to predict the sample harvest number for each required number of days.

(S20) The iteration control unit 128 determines whether all the training data records have been selected in step S18. If all the records have been selected, the process proceeds to step S21, and if there are unselected records in the training data, the process returns to step S18.

(S21) The iteration control unit 128 shifts the number of samples harvested on different fruiting days according to the harvesting dates so that the harvesting dates are aligned, and tallies the sample harvesting numbers for each harvesting date.
(S22) The iteration control unit 128 multiplies the total number of harvested samples for each harvest date by the reciprocal of the sample ratio to predict the total number of harvests for each harvest date.

(S23) The iteration control unit 128 compares the actual number of harvests for each harvest date indicated by the total data with the predicted number of harvests for each harvest date calculated in step S22, and calculates the total error. The total error is, for example, the residual sum of squares.

(S24) The iteration control unit 128 determines whether the iteration for evaluating the model error in steps S14 to S17 is the second time or later. If it is the second iteration or later, the process proceeds to step S25, and if it is the first iteration, the process proceeds to step S26.

(S25) The iteration control unit 128 determines whether the current total error is greater than the previous total error. If the current total error is greater than the previous total error, the process proceeds to step S26, and if the current total error is less than or equal to the previous total error, the process proceeds to step S27.

(S26) The iteration control unit 128 instructs the machine learning unit 127 to continue the iteration. The machine learning unit 127 saves the coefficients of the current prediction model, and updates the coefficients so that the model error calculated in step S17 becomes smaller. Then, the process returns to step S14.

(S27) The iteration control unit 128 reads the coefficients of the saved prediction model, and outputs the prediction model used in the previous iteration to the prediction model storage unit 124.
FIG. 16 is a flow chart showing an example of a harvest prediction procedure.

(S30) The data collection unit 125 collects the current year's weather data and sample data. The data collection unit 125 also collects weather forecast data after the forecast date.
(S31) The data processing unit 126 extracts the fruit-bearing date from the sample data, and classifies the sample fruit according to the fruit-bearing date. The data processing unit 126 counts the number of samples bearing fruits for each bearing day.

(S32) The data processing unit 126 extracts forecast temperature and forecast solar radiation from the forecast date to the harvest time from the weather forecast data. The data processing unit 126 converts the predicted temperature and amount of solar radiation into a predicted indoor temperature and amount of solar radiation using the environmental parameters.

(S33) The data processing unit 126 extracts the measured temperature and the measured amount of solar radiation from the date of fruiting to the day before the forecast date from the weather data for each day of fruiting. The data processing unit 126 sorts the measured temperatures and predicted temperatures by time, and calculates the average temperature at each time. In addition, the data processing unit 126 classifies the measured amount of solar radiation and the expected amount of solar radiation by time, and calculates the average amount of solar radiation at each time.

(S34) The harvest prediction unit 129 selects one fruit bearing date.
(S35) The harvest prediction unit 129 inputs the average temperature and the average amount of solar radiation corresponding to the selected fruit-bearing date to the prediction model, and predicts the required number of days distribution.

(S36) The harvest prediction unit 129 multiplies the sample number of fruit bearing corresponding to the selected fruit bearing date by the probability indicated by the required number of days distribution, and converts it into the number of sample harvests for each required number of days.
(S37) The harvest prediction unit 129 determines whether or not all fruit-bearing dates appearing in the sample data have been selected in step S34. If all dates have been selected, the process proceeds to step S38, and if there are unselected dates, the process returns to step S34.

(S38) The harvest prediction unit 129 shifts the sample harvest numbers on different fruit-bearing days according to the fruit-bearing days so that the harvest dates are aligned, and tallies the sample harvest numbers for each harvest date.
(S39) The harvest prediction unit 129 multiplies the total number of sample harvests for each harvest date by the reciprocal of the sample ratio to predict the total number of harvests for each harvest date.

(S40) The harvest prediction unit 129 outputs a prediction result indicating the number of harvests for each harvest date. For example, the harvest prediction unit 129 displays the prediction result on the display device 111 .
According to the information processing system of the second embodiment, using training data that associates the number of days required from fruit setting to harvest in the past year with the average temperature and average solar radiation during that period, the average temperature and average solar radiation A prediction model is learned that predicts the number of days required from the quantity. Then, the harvest date and number of harvests for the current year are predicted from the learned prediction model and the fruiting situation for the current year. Therefore, it is possible to provide useful information for farm management before the paprika is harvested.

Also, the prediction model is trained to output the probability distribution of the number of required days instead of the expected value of the number of required days. Therefore, it is possible to predict the variation in harvest dates, taking into consideration the individual differences in paprika that the growth rate varies greatly even when grown under the same growing environment. In addition, using the prediction model in the middle of learning, the total number of harvests predicted from the training data and the actual total number of harvests in the past year are compared, and when it is detected that the total error is minimized, the machine learning iteration is started. It is stopped and the prediction model when the total error is minimized is output. Therefore, it is possible to prevent the prediction model from outputting a probability distribution with excessively small variance due to over-learning, and it is possible to improve the prediction accuracy of the prediction model. Moreover, even from a small amount of sample data, a prediction model that appropriately reflects variations in harvest dates is generated. Therefore, it is possible to reduce the burden on the farmer who observes the sample fruit and collects the sample data.

REFERENCE SIGNS LIST 10 machine learning device 11 storage unit 12 processing unit 13 training data 14 total data 15 learning process 16 prediction model 17 prediction distribution

Claims (7)

to the computer,
training data each including a plurality of records in which information on the growing environment of the sample crop and the required number of days from the reference date when a predetermined state was observed to the harvest date of the sample crop are associated; and a plurality of records indicated by the plurality of records Obtain total data showing the actual distribution of the number of harvests for the harvest date for the crop set including the sample crops and other crops,
Generating a prediction model for calculating the probability distribution of the required number of days from information on the training environment, and using the training data to evaluate the error of the probability distribution calculated by the prediction model and updating the prediction model. start the iterative learning process,
In the middle of the learning process, a plurality of probability distributions calculated by the prediction model are synthesized from the information of the growing environment indicated by the plurality of records to calculate a predicted distribution of the number of harvests for the harvest date, and the predicted distribution is calculated. Determining when to stop the learning process based on the similarity between the actual distribution indicated by the total data,
A machine learning program that makes you do things. each of the sample crop and the other crop is a fruit;
The reference date is a date on which fruiting was observed,
The information on the growing environment of the sample crop includes the temperature and the amount of solar radiation from the date of fruiting to the date of harvest of the sample crop.
The machine learning program according to claim 1. In determining the stop timing, the similarity is evaluated each time the prediction model is updated, and when a peak of the similarity is detected, the learning process is stopped, and the learning process corresponding to the peak of the similarity is detected. Output the prediction model as a learning result,
3. The machine learning program according to claim 1 or 2. each of the plurality of records includes the reference date;
In determining the stop timing, combining the plurality of probability distributions corresponding to the plurality of records based on the reference date to calculate a sample prediction distribution of the number of harvests for the harvest date for the plurality of sample crops; calculating the predicted distribution for the crop set from the sample ratio of the plurality of sample crops to the crop set and the sample predicted distribution;
The machine learning program according to any one of claims 1 to 3. In the determination of the stop timing, the learning process is stopped when the similarity indicates that the similarity is equal to or greater than a predetermined standard.
The machine learning program according to any one of claims 1 to 4. the computer
training data each including a plurality of records in which information on the growing environment of the sample crop and the required number of days from the reference date when a predetermined state was observed to the harvest date of the sample crop are associated; and a plurality of records indicated by the plurality of records Obtain total data showing the actual distribution of the number of harvests for the harvest date for the crop set including the sample crops and other crops,
Generating a prediction model for calculating the probability distribution of the required number of days from information on the training environment, and using the training data to evaluate the error of the probability distribution calculated by the prediction model and updating the prediction model. start the iterative learning process,
In the middle of the learning process, a plurality of probability distributions calculated by the prediction model are synthesized from the information of the growing environment indicated by the plurality of records to calculate a predicted distribution of the number of harvests for the harvest date, and the predicted distribution is calculated. Determining when to stop the learning process based on the similarity between the actual distribution indicated by the total data,
machine learning method. training data each including a plurality of records in which information on the growing environment of the sample crop and the required number of days from the reference date when a predetermined state was observed to the harvest date of the sample crop are associated; and a plurality of records indicated by the plurality of records a storage unit for storing total data indicating the actual distribution of the number of harvests with respect to the harvest date for a set of crops including sample crops and other crops;
Generating a prediction model for calculating the probability distribution of the required number of days from information on the training environment, and using the training data to evaluate the error of the probability distribution calculated by the prediction model and updating the prediction model. Repeated learning processing is started, and in the middle of the learning processing, a plurality of probability distributions calculated by the prediction model are synthesized from the information of the growing environment indicated by the plurality of records to obtain a prediction distribution of the number of harvests for the harvest date. a processing unit that calculates and determines when to stop the learning process based on the degree of similarity between the predicted distribution and the actual distribution indicated by the total data;
A machine learning device having

Images