Prediction device and prediction method

Abstract

A prediction device 1 comprises: a generation unit 22 that refers to priority data 11 associating each of parameters used in a prediction model for predicting the concentration of each element of a marine ecosystem model with a high priority or a low priority, and generates a plurality of high priority sets 12 including any of a plurality of values identified for each parameter of the high priority; a prediction unit 23 that inputs each of a plurality of parameter sets including any of the plurality of high priority sets 12 and a predetermined value of each parameter of the low priority to the prediction model, and outputs a plurality of result sets 15 including a value predicted for the concentration of each element; and a selection unit 24 that selects, from among the plurality of result sets, the high priority set 12 used upon predicting a result set having a small difference from an actual measured value of the concentration of each element.

Description

Prediction device and prediction method

[0001] The present disclosure relates to a prediction device and a prediction method.

[0002] There is an NPZD model for predicting the amount of substances or biomass in a marine ecosystem (see Non-Patent Document 1). The NPZD model formulates four elements: nutrients (N: Nutrient), phytoplankton (P: Phytoplankton), zooplankton (Z: Zooplankton), and detritus (D: Detritus).

[0003] In order to accurately predict the concentration of each element using a marine ecosystem model, parameter tuning is performed. Parameter tuning repeats the adjustment of parameter values used in the marine ecosystem model and the evaluation of the model using a grid search method or the like. The grid search method obtains the optimal parameters after executing the model with all combinations of possible parameter values (Non-Patent Document 2).

[0004] As an optimization method for efficiently obtaining the optimal solution, there is Bayesian optimization (Non-Patent Document 3). Bayesian optimization iteratively searches for candidate points with a high probability of obtaining the optimal solution from limited actual data for a black box function.

[0005] “A Practical Guide to Ecological Modelling”, Karline Soetaert and Peter MJ Herman, 2009.Liashchynskyi, Petro, and Pavlo Liashchynskyi. “Grid search, random search, genetic algorithm: a big comparison for NAS.”, 2019Brochu, Eric, Vlad M. Cora, and Nando De Freitas. “A tutorial on Bayesian optimization of expensive cost functions, with application to active user modeling and Hierarchical reinforcement learning.", 2010

[0006] Grid search methods run the model with all possible combinations of parameter values, which places a heavy burden on parameter tuning.

[0007] This disclosure is made in view of the above circumstances, and its purpose is to provide a technology that can reduce the burden of parameter tuning in marine ecosystem models.

[0008] A prediction device according to one aspect of the present disclosure includes: a generation unit that generates a plurality of high-priority sets containing one of a plurality of values ​​identified for each high-priority parameter by referring to priority data that associates high priority or low priority with each parameter used in a prediction model that predicts the concentration of each element of a marine ecosystem model; a prediction unit that inputs one of the plurality of high-priority sets and each of a plurality of parameter sets containing predetermined values ​​for each low-priority parameter to the prediction model and outputs a plurality of result sets containing predicted values ​​for the concentration of each element; and a selection unit that selects the high-priority set used when predicting a result set in which the difference from the measured value of the concentration of each element is small.

[0009] A prediction method in one aspect of the present disclosure involves a computer referencing priority data that associates high priority or low priority with each parameter used in a prediction model for predicting the concentration of each element of a marine ecosystem model, generating a plurality of high-priority sets containing one of a plurality of values ​​identified for each high-priority parameter, inputting one of the plurality of high-priority sets and each of a plurality of parameter sets containing predetermined values ​​for each low-priority parameter into the prediction model, outputting a plurality of result sets containing predicted values ​​for the concentration of each element, and selecting from the plurality of result sets the result set in which the difference from the measured value of the concentration of each element is small, which is the high-priority set used when making the prediction.

[0010] This disclosure provides a technology that can reduce the burden of parameter tuning in marine ecosystem models.

[0011] Figure 1 is a diagram illustrating the functional blocks of the prediction device relating to this disclosure. Figure 2 is a diagram illustrating an example of a marine ecosystem model. Figure 3 is a diagram illustrating an example of a mathematical formula for identifying a marine ecosystem model. Figure 4 is a diagram illustrating the data structure and an example of the data for priority data. Figure 5 is a diagram illustrating the data structure and an example of the data for multiple high-priority sets. Figure 6 is a diagram illustrating the data structure and an example of the data for multiple low-priority sets. Figure 7 is a diagram illustrating the data structure and an example of the data for multiple result sets. Figure 8 is a diagram illustrating the data structure and an example of the data for parameter data. Figure 9 is a sequence diagram illustrating the prediction method relating to this disclosure (part 1). Figure 10 is a sequence diagram illustrating the prediction method relating to this disclosure (part 2). Figure 11 is a diagram illustrating the hardware configuration of the computer used in the prediction device.

[0012] Embodiments of this disclosure will be described below with reference to the drawings. In the drawings, the same parts are denoted by the same reference numerals and their descriptions are omitted.

[0013] The prediction device 1 shown in Figure 1 efficiently improves the accuracy of prediction models in marine ecosystem models. The prediction device 1 efficiently tunes the parameters used in prediction models that predict the concentrations of each element in a marine ecosystem model.

[0014] Specifically, the prediction device 1 first tunes the high-priority parameters among the parameters used in the prediction model, and then tunes the low-priority parameters.

[0015] When tuning high-priority parameters, the prediction device 1 generates multiple parameter sets containing multiple values ​​for each high-priority parameter. In each of the multiple parameter sets, multiple different values ​​are set for each high-priority parameter, and a predetermined value (fixed value) is set for each low-priority parameter. The prediction device 1 uses the multiple parameter sets to predict the concentration of each element in the marine ecosystem model. The prediction device 1 evaluates the prediction results, and if it determines that the predetermined accuracy cannot be obtained, it generates one or more new parameter sets using a Bayesian optimization method. The prediction device 1 repeats the process of predicting the concentration of each element for each of the new parameter sets and evaluating the prediction results to tune each high-priority parameter.

[0016] When tuning low-priority parameters, the prediction device 1 generates multiple parameter sets containing multiple values ​​for each low-priority parameter. Each of the multiple parameter sets is assigned multiple different values ​​for each low-priority parameter, and the tuned values ​​for each high-priority parameter are assigned. The prediction device 1 uses the multiple parameter sets to predict the concentration of each element in the marine ecosystem model. The prediction device 1 evaluates the prediction results, and if it determines that the required accuracy cannot be obtained, it generates one or more new parameter sets using a Bayesian optimization method. The prediction device 1 repeats the process of predicting the concentration of each element for each of the new parameter sets and evaluating the prediction results to tune each low-priority parameter.

[0017] This allows the prediction device 1 to efficiently improve the accuracy of the prediction model.

[0018] Generally, marine ecosystem models have a large number of parameters that need to be tuned. In evaluating these parameters, multiple simulations are repeated, with each parameter being assigned one of several candidate values.

[0019] The more parameters explored at once, the longer the computation time required for the search. The prediction device 1 according to this disclosure divides multiple parameters into multiple groups and performs stepwise tuning for each group, thereby suppressing computation time.

[0020] When a large number of parameters are explored at once, Bayesian optimization may result in a bias in the parameters being optimized. For example, as the number of iterations of the search progresses, there is a tendency to concentrate on optimizing certain parameters while neglecting the optimization of other parameters. Therefore, the prediction device 1 according to this disclosure can prevent a bias in the parameters being optimized by performing the parameter search in stages, thereby reducing the number of parameters optimized in each stage.

[0021] The prediction target of prediction device 1 is, for example, a marine ecosystem model as shown in Figure 2. Figure 2 shows the activity of each element in a tank: nutrients (N), phytoplankton (P), zooplankton (Z), and detritus (D). The concentration of each element increases or decreases due to the activities of each element, such as uptake, respiration, predation, decomposition, and death.

[0022] Note that the prediction model shown in Figure 2-3 is just one example and is not limited to this model.

[0023] The increase or decrease of each element in the environment shown in Figure 2 is shown in the formula in Figure 3. The prediction device 1 according to this disclosure calculates the predicted concentration of each element over time using the parameter values ​​used in the prediction model and the formula shown in Figure 3. The prediction device 1 evaluates the difference between the predicted value and the measured value and tunes the parameters.

[0024] As shown in Figure 1, the prediction device 1 includes priority data 11, high-priority set 12, low-priority set 13, prediction model data 14, result set 15, measured value data 16, and parameter data 17, as well as the functions of a setting unit 21, a generation unit 22, a prediction unit 23, and a selection unit 24. Each data, and data such as initial values ​​necessary for tuning the parameters, are stored in a storage device such as a memory 902 or storage 903. Each function is implemented in the CPU 901.

[0025] Priority data 11 associates high priority or low priority with each parameter used in the prediction model that forecasts the concentration of each element in the marine ecosystem model. As shown in Figure 4, priority data 11 associates high priority or low priority with each parameter used in the prediction model. Priority data 11 further associates upper and lower limits.

[0026] The upper and lower limits specify the range of values ​​each parameter can take. The upper and lower limits are examples of indicators for classifying each parameter as high-priority or low-priority. Priority data 11 may further associate indicators for classifying each parameter as high-priority or low-priority.

[0027] This disclosure describes, but is not limited to, a case in which parameters are divided into two groups and tuned in two stages. The prediction device 1 may divide the parameters into three or more groups and tune the parameters of each group in stages.

[0028] The high-priority set 12 contains one of several values ​​identified for each high-priority parameter. As shown in Figure 5, the prediction device 1 stores multiple high-priority sets 12. The multiple high-priority sets 12 have different combinations of values ​​for each high-priority parameter.

[0029] The low-priority set 13 contains one of several values ​​identified for each low-priority parameter. As shown in Figure 6, the prediction device 1 stores multiple low-priority sets 13. The multiple low-priority sets 13 have different combinations of values ​​for each low-priority parameter.

[0030] Predictive model data 14 is data that identifies the predictive model in the marine ecosystem model. Predictive model data 14 is, for example, the equation in Figure 3.

[0031] The result set 15 is data resulting from the prediction of the concentration of each element over time by the prediction unit 23. One result set 15 is generated for each parameter set. The result set 15 associates the prediction result with the identifier of the parameter set input when obtaining that prediction result.

[0032] As shown in Figure 7, the prediction device 1 stores multiple result sets 15. In Figure 7, the predicted concentration of each element over time is associated with each of the multiple high-priority sets 12. The identifier of the high-priority set 12 identifies the parameter set. As will be described later, instead of the identifier of the high-priority set, the identifier of the low-priority set 13 or the identifier of the parameter set input from the selection unit 24 may be set. The parameter set input from the selection unit 24 is a parameter set that includes the values ​​of the high-priority parameters selected by the Bayesian optimization method, or a parameter set that includes the values ​​of the low-priority parameters selected by the Bayesian optimization method.

[0033] The measured data 16 consists of data of values ​​actually observed in the environment that realizes the environment targeted for prediction by the prediction device 1. The measured data 16 is data that correlates the concentration of each element over time in the realized environment.

[0034] Parameter data 17 consists of the values ​​of the parameters referenced during prediction. Parameter data 17 associates each parameter defined by priority data 11 with its corresponding value. Parameter data 17 transitions depending on the input parameter set.

[0035] Figure 8(a) shows the parameter data 17 that is referenced when tuning high-priority parameters. No values ​​are set for high-priority parameters. Fixed values ​​are set for low-priority parameters.

[0036] Figure 8(b) shows the parameter data 17 referenced when tuning low-priority parameters. The values ​​selected through tuning are set for the high-priority parameters. No values ​​are set for the low-priority parameters.

[0037] Figure 8(c) shows the parameter data 17 that is referenced when simulating according to predetermined prediction conditions after tuning of each parameter is complete. The values ​​selected through tuning are set for the high-priority and low-priority parameters.

[0038] The setting unit 21 sets a priority for each parameter of the priority data 11. The setting unit 21 generates the priority data 11 by associating high priority or low priority for each parameter from the search range of each parameter used in the prediction model.

[0039] The setting unit 21 sets priorities, for example, based on the difficulty in selecting the optimal value for a parameter. In this disclosure, the setting unit 21 sets a high priority for parameters where the difference between the upper and lower limits of the parameter is greater than a predetermined value. The setting unit 21 sets a low priority for parameters where the difference is smaller than a predetermined value. This predetermined value is set, for example, as a percentage of a predetermined reference value, such as 10% of the upper or lower limit of the parameter.

[0040] The generation unit 22 refers to the priority data 11 and generates multiple high-priority sets 12, each containing one of the multiple values ​​identified for each high-priority parameter. The generation unit 22 identifies multiple values ​​between the upper and lower limits for each high-priority parameter in the priority data 11. For example, the generation unit 22 randomly selects multiple values ​​for each parameter. The generation unit 22 generates multiple high-priority sets 12, each containing one of the multiple values ​​identified for each parameter.

[0041] The generation unit 22 generates multiple low-priority sets 13, each containing one of the multiple values ​​identified for each low-priority parameter. The generation unit 22 generates the multiple low-priority sets 13 in the same way as the high-priority sets 12.

[0042] The prediction unit 23 uses the input parameter set and the prediction model data 14 to predict the concentration of each element and outputs a result set 15. The prediction unit 23 outputs the result set 15 for each parameter set.

[0043] For tuning of high-priority parameters, the prediction unit 23 sets one of the plurality of high-priority sets 12 in the parameter set and predicts the concentration of each element. When the tuning of the high-priority parameters is completed, the prediction unit 23 sets one of the plurality of low-priority sets 13 in the parameter set for tuning of the low-priority parameters and predicts the concentration of each element. When the tuning of each parameter is completed, the prediction unit 23 uses the values of the tuned parameters to specify arbitrary prediction conditions and predicts the concentration of each element.

[0044] The prediction process for tuning of high-priority parameters will be described. The prediction unit 23 acquires a plurality of parameter sets for tuning of high-priority parameters. One parameter set includes one of the plurality of high-priority sets 12 and predetermined values of each low-priority parameter. The predetermined value of each low-priority parameter is, for example, a fixed value acquired from the parameter data 17.

[0045] The prediction unit 23 inputs each of the plurality of parameter sets into the prediction model and outputs a plurality of result sets 15 including predicted values for the concentration of each element. The prediction unit 23 outputs the result set 15 in association with the identifier of the high-priority set 12 used in the prediction and its prediction result. From the plurality of result sets 15, the selection unit 24 selects the high-priority set 12 and the tuning of the high-priority parameters is completed.

[0046] The prediction process for tuning of low-priority parameters will be described. The prediction unit 23 acquires a plurality of parameter sets for tuning of low-priority parameters. One parameter set includes the high-priority set​​The prediction unit 23 inputs each of a plurality of parameter sets for tuning low-priority parameters into the prediction model, and outputs a plurality of result sets 15 including predicted values for the concentrations of each element. The prediction unit 23 outputs the result set 15 in association with the identifier of the low-priority set 13 used when making the prediction and its prediction result. From the plurality of result sets 15, the selection unit 24 selects the low-priority set 13, and the tuning of the low-priority parameters ends.

[0048] In the present disclosure, when the high-priority set 12 or the low-priority set 13 generated by the generation unit 22 does not satisfy a predetermined condition, specifically, when a predetermined accuracy is not achieved, as will be described later, a new parameter set (another parameter set) is input from the selection unit 24. The prediction unit 23 inputs each of the new parameter sets input from the selection unit 24 into the prediction model, and outputs a plurality of result sets 15 including predicted values for the concentrations of each element. The prediction unit 23 outputs the result set 15 in association with the identifier of the parameter set used when making the prediction and its prediction result. The prediction unit 23 repeats the prediction until a predetermined end condition is satisfied. The predetermined end condition is that a predetermined accuracy is achieved with a new parameter set, the number of prediction times, or the processing time, etc.

[0049] When the tuning of each parameter ends, the prediction unit 23 inputs a parameter set including the values of the parameters of the high-priority set 12 and the low-priority set 13 selected by the selection unit 24 and arbitrary prediction conditions into the prediction model, and outputs a result set 15 including the predicted values for the concentrations of each element.

[0050] The selection unit 24 selects suitable parameters by referring to the plurality of result sets 15 output by the prediction unit 23. In the present disclosure, the selection unit 24 selects high-priority parameters and low-priority parameters.

[0051] Among the plurality of result sets 15 output by the prediction process for tuning the high-priority parameters, the selection unit 24 selects the high-priority set 12 used when predicting a result set with a small difference from the actually measured value of the concentration of each element.

[0052] The selection unit 24 calculates an evaluation value for each of the multiple result sets 15 output by the prediction process for tuning high-priority parameters, which evaluates the difference between the actual measured values ​​of the concentrations of each element and the result set 15.

[0053] Here, the evaluation value assesses the difference between the predicted value output by the prediction model and the actual value. The evaluation value is calculated from the RMSE (Root Mean Squared Error) shown in equation (1). A smaller evaluation value means that the predicted value is closer to the actual value, and a larger evaluation value means that the predicted value is further away from the actual value.

[0054]

[0055] The selection unit 24 selects the high-priority set 12 corresponding to the lowest evaluation value if the lowest evaluation value is below the threshold. The selection unit 24 sets the selected high-priority set 12 in the parameter data 17. If the lowest evaluation value is above the threshold, the selection unit 24 determines that none of the multiple high-priority sets 12 generated by the generation unit 22 will lead to the desired accuracy. The selection unit 24 inputs a new parameter set, including the values ​​of each high-priority parameter selected by the Bayesian optimization method, to the prediction unit 23.

[0056] The selection unit 24 inputs a parameter set derived from at least one of a plurality of result sets 15 associated with each of the high-priority parameter sets 12, along with the evaluation values ​​of that result set 15, into a module that implements a Bayesian optimization method. This module uses the parameter set and its evaluation values ​​to obtain a posterior distribution using Gaussian process regression. The module uses an acquisition function to select one or more new high-priority parameter sets from the obtained posterior distribution. Each high-priority parameter set includes the values ​​of each high-priority parameter. The selection unit 24 generates one or more new parameter sets. Each new parameter set includes the values ​​of the high-priority parameters selected by the module and predetermined values ​​of the low-priority parameters. The selection unit 24 inputs one or more new parameter sets into the prediction unit 23.

[0057] From among the multiple result sets 15 output by the prediction process for tuning low-priority parameters, the selection unit 24 selects the low-priority set 13 used when predicting the result set with the smallest difference between the predicted result set and the measured values ​​of the concentrations of each element. The selection unit 24 sets the selected low-priority set 13 into the parameter data 17.

[0058] The selection unit 24 calculates an evaluation value for each of the multiple result sets 15 output by the prediction process for tuning low-priority parameters, evaluating the difference between the measured values ​​and the evaluation value of each element. The evaluation value is calculated, for example, from equation (1).

[0059] The selection unit 24 selects the lowest priority set 13 corresponding to the lowest evaluation value if the lowest evaluation value is below the threshold. The selection unit 24 sets the selected low priority set 13 in the parameter data 17. If the lowest evaluation value is above the threshold, the selection unit 24 determines that none of the multiple low priority sets 13 generated by the generation unit 22 will lead to the desired accuracy. The selection unit 24 inputs a new parameter set, including the values ​​of each low priority parameter selected by the Bayesian optimization method, to the prediction unit 23.

[0060] The selection unit 24 inputs the parameter set derived from at least one result set 15 from among a plurality of result sets 15 associated with each of the low-priority parameter sets 13, and the evaluation values ​​of the result set 15, into a module that implements a Bayesian optimization method. This module uses the parameter set and its evaluation values ​​to obtain a posterior distribution using Gaussian process regression. The module uses an acquisition function to select one or more new low-priority parameter sets from the obtained posterior distribution. The low-priority parameter set includes the values ​​of each low-priority parameter. The selection unit 24 generates one or more new parameter sets. The new parameter sets include the values ​​of the low-priority parameters selected by the module and the tuned values ​​of the high-priority parameters. The selection unit 24 inputs one or more new parameter sets into the prediction unit 23.

[0061] The selection unit 24 repeatedly generates new parameters and has the prediction unit 23 make predictions until a predetermined termination condition is met. The predetermined termination condition is the number of predictions or processing time at which a predetermined accuracy is achieved with the new parameter set.

[0062] The evaluation value shown in equation (1) is calculated to be lower the smaller the difference between the measured value and the predicted value. Therefore, we will explain the case where the parameter set used to calculate the predicted value is adopted when the evaluation value is below the threshold. The adoption criteria may be changed depending on the method for calculating the evaluation value.

[0063] (Prediction Method) The prediction method according to this disclosure will be described with reference to Figures 9 and 10. The processes in Figures 9 and 10 are executed, for example, when an instruction to tune the parameters in the prediction model is input from a user terminal.

[0064] In step S11, the generation unit 22 refers to the priority data 11 and generates a plurality of high-priority sets 12, each containing the value of a high-priority parameter.

[0065] In step S12, the prediction unit 23 predicts the concentration of each element over time using each of the input parameter sets. Here, the prediction unit 23 makes the prediction using each of the high-priority sets 12 generated by the generation unit 22 in step S11.

[0066] The prediction unit 23 uses the values ​​set in the high-priority set 12 for each of the high-priority parameters. The prediction unit 23 uses predetermined values ​​set in the parameter data 17 for each of the low-priority parameters. The prediction unit 23 stores the predicted results in the result set 15. The result set 15 is generated for each of the multiple high-priority sets 12 generated by the generation unit 22 in step S11.

[0067] In step S13, the selection unit 24 compares the multiple result sets 15 with the measured value data 16 to determine whether there is a result set 15 that satisfies a predetermined accuracy. The selection unit 24 calculates an evaluation value for each result set 15. If there is an evaluation value among the calculated evaluation values ​​that satisfies the predetermined accuracy, the selection unit 24 selects a high-priority set 12 corresponding to the evaluation value that satisfies the predetermined accuracy. The selection unit 24 sets each value of the selected high-priority set 12 into the parameter data 17.

[0068] If the selection unit 24 determines in step S13 that the result set 15 does not meet a predetermined accuracy, in step S14 the selection unit 24 generates one or more new parameter sets using a Bayesian optimization method. The selection unit 24 inputs the generated new parameter sets to the prediction unit 23. In step S12 the prediction unit 23 makes predictions using each of the new parameter sets input from the selection unit 24. The prediction device 1 may repeat the process from step S12 to step S14 until a predetermined termination condition is met.

[0069] The tuning of high-priority parameters is completed through the processing in steps S11 to S14.

[0070] In step S21, the generation unit 22 refers to the priority data 11 and generates a plurality of low-priority sets 13, each containing the value of a low-priority parameter.

[0071] In step S22, the prediction unit 23 predicts the concentration of each element for each time period using each of the input parameter sets. Here, the prediction unit 23 makes predictions using each of the low-priority sets 13 generated by the generation unit 22 in step S21.

[0072] The prediction unit 23 uses the values ​​set in the low-priority set 13 for each of the low-priority parameters. The prediction unit 23 uses the values ​​set in the parameter data 17, specifically the high-priority set 12 selected in the process shown in Figure 9, for each of the high-priority parameters. The prediction unit 23 stores the predicted results in the result set 15. The result set 15 is generated for each of the multiple low-priority sets 13 generated by the generation unit 22 in step S21.

[0073] In step S23, the selection unit 24 compares the multiple result sets 15 with the measured value data 16 to determine whether there is a result set 15 that satisfies a predetermined accuracy. The selection unit 24 calculates an evaluation value for each result set 15. If there is an evaluation value among the calculated evaluation values ​​that satisfies the predetermined accuracy, the selection unit 24 selects a low-priority set 13 corresponding to the evaluation value that satisfies the predetermined accuracy. The selection unit 24 sets each value of the selected low-priority set 13 into the parameter data 17.

[0074] In step S23, if the selection unit 24 determines that the result set 15 does not meet a predetermined accuracy, in step S24, the selection unit 24 generates one or more new parameter sets using a Bayesian optimization method. The selection unit 24 inputs the generated new parameter sets to the prediction unit 23. In step S22, the prediction unit 23 makes predictions using each of the new parameter sets input from the selection unit 24. The prediction device 1 may repeat the process from step S22 to step S24 until a predetermined termination condition is met.

[0075] The processing in steps S21 to S24 completes the tuning of the low-priority parameters. In step S23, once the tuning of the low-priority parameters is complete, the tuned values ​​for each parameter are set in the parameter data 17.

[0076] In step S25, the prediction unit 23 uses the tuned values ​​of each parameter to predict the concentration of each element under predetermined prediction conditions. The prediction unit 23 outputs the predicted results to the result set 15.

[0077] Such a prediction device 1 can reduce the burden of parameter tuning in marine ecosystem models. Combined with the fact that the prediction device 1 employs Bayesian optimization methods, it can easily construct highly accurate models.

[0078] This makes it possible to conserve biodiversity, for example, by optimizing human activities to predict that biomass in marine ecosystems will not decrease.

[0079] Furthermore, by applying the tuning method described herein to land-based aquaculture, which artificially constitutes a marine ecosystem on land, for example, it is expected that benefits can be expected from maximizing aquaculture production. In land-based aquaculture environments, the amount of nutrients added and the amount of circulating water can be optimized relative to the required production volume, and material costs can be expected to be reduced. In land-based aquaculture environments, the need for manual adjustment of additives during operation can be eliminated, and operating costs can be reduced. The development of the land-based aquaculture industry can contribute to reducing environmental burdens, such as the utilization of aquaculture water. Ultimately, land-based aquaculture could become a part of a closed ecosystem life support system in a space station or space base.

[0080] The prediction device 1 described above in this disclosure uses, for example, a general-purpose computer system comprising a CPU (Central Processing Unit, processor) 901, memory 902, storage 903 (HDD: Hard Disk Drive, SSD: Solid State Drive), communication device 904, input device 905, and output device 906. In this computer system, each function of the prediction device 1 is realized by the CPU 901 executing a program loaded onto the memory 902.

[0081] The prediction device 1 may be implemented on a single computer, or on multiple computers. Furthermore, the prediction device 1 may be a virtual machine implemented on a computer.

[0082] The program of prediction device 1 can be stored on a computer-readable recording medium such as an HDD, SSD, USB (Universal Serial Bus) memory, CD (Compact Disc), or DVD (Digital Versatile Disc), or it can be distributed over a network. A computer-readable recording medium is, for example, a non-transitory recording medium.

[0083] This disclosure is not limited to the embodiments described above, and numerous modifications are possible within the scope of its essence.

[0084] 1 Prediction device 11 Priority data 12 High priority set 13 Low priority set 14 Prediction model data 15 Result set 16 Measured value data 17 Parameter data 21 Setting unit 22 Generation unit 23 Prediction unit 24 Selection unit 901 CPU 902 Memory 903 Storage 904 Communication device 905 Input device 906 Output device

Claims

1. A prediction device comprising: a generation unit that generates multiple high-priority sets containing one of multiple values ​​identified for each high-priority parameter by referring to priority data that associates high priority or low priority with each parameter used in a prediction model for predicting the concentration of each element of a marine ecosystem model; a prediction unit that inputs one of the multiple high-priority sets and each of the multiple parameter sets containing predetermined values ​​for each low-priority parameter into the prediction model and outputs multiple result sets containing predicted values ​​for the concentration of each element; and a selection unit that selects the high-priority set used when predicting a result set in which the difference from the measured value of the concentration of each element is small.

2. The prediction device according to claim 1, wherein the generation unit further generates a plurality of low-priority sets, each containing one of a plurality of values ​​identified for each of the low-priority parameters; the prediction unit further inputs each of a plurality of other parameter sets, each containing the selected high-priority set and one of the plurality of low-priority sets, into the prediction model and outputs a plurality of other result sets containing predicted values ​​for the concentration of each element; and the selection unit selects from the plurality of other result sets the low-priority set used when predicting the other result set in which the difference from the measured value of the concentration of each element is small.

3. The prediction device according to claim 2, wherein the prediction unit inputs a parameter set including the values ​​of each parameter in the selected high-priority set and low-priority set to the prediction model and outputs a result set including predicted values ​​for the concentration of each element.

4. The prediction device according to claim 1, wherein the selection unit calculates an evaluation value for each of the plurality of result sets, which evaluates the difference between the measured value and the concentration of each element, and if the evaluation value does not satisfy a predetermined condition, the selection unit repeats the process of inputting a parameter set including the values ​​of each high-priority parameter selected by the Bayesian optimization method to the prediction unit.

5. The prediction device according to claim 4, wherein the selection unit obtains the values ​​of each high-priority parameter selected by the Bayesian optimization method from at least one result set among the plurality of result sets and the evaluation values ​​of the result set.

6. The prediction device according to claim 1, further comprising a setting unit that generates priority data by associating high priority or low priority of each parameter with the search range of each parameter used in the prediction model.

7. A prediction method comprising: a computer generating multiple high-priority sets containing one of multiple values ​​identified for each high-priority parameter by referring to priority data that associates high priority or low priority with each parameter used in a prediction model for predicting the concentration of each element of a marine ecosystem model; inputting one of the multiple high-priority sets and each of the multiple parameter sets containing predetermined values ​​for each low-priority parameter into the prediction model to output multiple result sets containing predicted values ​​for the concentration of each element; and selecting from the multiple result sets the result set in which the difference from the measured value of the concentration of each element is small, which is the high-priority set used when making the prediction.

Images