WO2022158042A1

WO2022158042A1 - Prediction system and prediction method

Info

Publication number: WO2022158042A1
Application number: PCT/JP2021/034357
Authority: WO
Inventors: 純一大崎; 好彦赤城; 太富澤; 隆岡田; 雄希奥田
Original assignee: 日立Astemo株式会社
Priority date: 2021-01-22
Filing date: 2021-09-17
Publication date: 2022-07-28
Also published as: JP2024038530A

Abstract

In the present invention, a microcomputer (101) (processor): stores, in a RAM (102) (first memory), a prediction result indicating a state at a second timing which is predicted from a state observed at a first timing using a plurality of prediction models; and generates, from a state observed at the second timing, a correct answer state indicating a correct answer with respect to the prediction result. The microcomputer (101) compares the correct answer state with the prediction result for each prediction model, calculates a correct answer rate for each prediction model, and calculates a prediction experience value for each prediction model indicating the number of predictions that were carried out for each prediction model. The microcomputer (101) selects and outputs one prediction result from the plurality of prediction results using the correction answer rate for each prediction model and the prediction experience value for each prediction model.

Description

Forecasting system and forecasting method

The present invention relates to a prediction system and a prediction method.

In the development of automobiles in recent years, in addition to fuel efficiency improvement technology, development that seeks safety and convenience such as preventive safety technology and automatic driving technology is being actively pursued. There is predictive control.

One example of vehicle behavior predictive control is control that estimates the positional relationship between the vehicle traveling ahead and the own vehicle. Based on the predicted future state, the driver will be able to accelerate or not after a few seconds. There is a judgment such as whether control to stop immediately should be performed.

　For these vehicle behavior prediction controls, safety is always required and high prediction accuracy is required, and methods to improve prediction accuracy include changing the calculation method that is the basis of prediction and changing constants. As another method, a system has been invented that predicts the future state of an object using a plurality of prediction models and outputs the prediction result of the most probable prediction model.

In Patent Document 1, in order to estimate the state of a secondary battery for electric vehicles, a plurality of models are made to make predictions using different prediction methods, and at the same time, the model with the smallest deviation between the actual state and the predicted state is evaluated. proposed a technique for evaluating the accuracy of each prediction model and outputting the result of the prediction model with the highest accuracy.

When prior patent document 1 is applied to vehicle behavior prediction, a plurality of prediction models with different calculation methods are prepared, the accuracy of these prediction models is evaluated, and the results of which prediction model to adopt are selected to determine the overall accuracy. expected to improve.

JP 2013-190274 A

However, prior patent document 1 assumes a secondary battery for electric vehicles as a target, and it takes time to compare the state estimated by each prediction model and the actual state and find out which prediction model has the highest accuracy. time consuming. For example, when predicting a state with a short prediction period of several milliseconds or seconds, there is a concern that the wrong prediction model will be selected immediately after the start of prediction, reducing the overall prediction accuracy.

In addition, when developing a forecast model based on a new way of thinking, while improvements were made in areas where forecast accuracy was low in the conventional forecast model, there was no impact on the forecast results in areas where the conventional forecast model maintained high forecast accuracy. It is desirable to be able to prevent

In prior patent document 1, the parameters used by the prediction models are optimized in order to improve the accuracy of each prediction model, but the prediction parameters are averaged between the parts that were successfully predicted by the prediction model and the parts that were not. may occur, resulting in a loss of prediction accuracy.

The present invention has been made in view of such problems, and its purpose is to predict the future situation based on a certain current situation based on the prediction results of a plurality of prediction models. An object of the present invention is to provide a prediction device (system) that enables highly accurate future situation prediction for the results of each prediction model.

In order to achieve the above object, one example of the present invention is a prediction device comprising a processor and a first memory, wherein the processor uses a plurality of prediction models at a first timing to obtain states observed at that time. storing in the first memory a prediction result indicating a state at a second timing predicted from the state observed at the second timing, generating a correct state indicating a correct state for the prediction result from the state observed at the time of the second timing; Comparing the correct state and the prediction result for each prediction model, calculating the correct answer rate for each prediction model, and calculating the prediction empirical value for each prediction model, which indicates the number of predictions performed for each prediction model. Then, using the correct answer rate for each prediction model and the prediction experience value for each prediction model, one prediction result is selected from a plurality of prediction results and output.

According to the present invention, highly accurate future situation prediction is possible. Problems, configurations, and effects other than those described above will be clarified by the following description of the embodiments.

It is a block diagram of the example which applied the system which concerns on this embodiment to the control apparatus mounted in a vehicle. FIG. 2 is an explanatory diagram of an application example of a driver acceleration/deceleration prediction application of the control device illustrated in FIG. 1 ; It is a block diagram in a first embodiment. FIG. 4 is an explanatory diagram of processing of a measurement result correct answer determination unit and an arbitration parameter storage unit; It is explanatory drawing of the example which calculates a prediction empirical value from the total of the number of correct answers and the number of incorrect answers. It is a figure explaining 2nd embodiment at the time of adding a prediction model with respect to 1st embodiment. It is a figure explaining the correct answer rate of each prediction model, and the change of prediction empirical value. It is a control-apparatus flowchart figure in 1st embodiment and 2nd embodiment. It is a block diagram for explaining a third embodiment to which the present invention is applied. A table of state IDs corresponding to states discriminated by the state discriminator. FIG. 11 is an explanatory diagram of an example of performing arbitration using a predicted percentage of correct answers using state-based arbitration parameters; It is the flowchart figure which showed the flow of the predictive control of the control apparatus in 3rd embodiment.

The configuration and operation of the system according to the first to third embodiments of the present invention will be described below with reference to the drawings. The present embodiment relates to a system for predicting a future state from a plurality of prediction models, and more particularly to technology for arbitrating the prediction results of the plurality of prediction models and determining the predicted state output by the system. Although the drawings specifically show an application example to a vehicle, they are drawings for explaining the basic principle of the present invention, and the scope of the present invention is not limited to the scope described in the drawings.

[First embodiment]
FIG. 1 shows an example in which the prediction system according to the first embodiment is applied to a control device (prediction device) mounted on a vehicle. In the present embodiment, prediction of acceleration/deceleration of a driver driving a vehicle will be described as an example. Although the vehicle and the control device are mentioned in the present embodiment, the scope of application of the present embodiment is not limited to the prediction target exemplified above and the vehicle and the control device exemplified as the application destination of the present embodiment.

The control device 100 includes a microcomputer 101 (CPU), RAM 102 (Random Access Memory), ROM 103 (Read Only Memory), and nonvolatile ROM 104 (EEPROM). Information obtained from the vehicle is input to the control device 100 .

In the control device 100, the distance to the preceding vehicle obtained based on the information from the camera sensor 105 attached to the vehicle, the vehicle speed obtained from the vehicle speed sensor 106, and the vehicle speed obtained from the acceleration sensor 107 are stored. The acceleration of the own vehicle and the accelerator opening obtained from the accelerator opening sensor 108 are input.

These inputs are input to the control device 100, and the control device 100 determines whether or not the driver will accelerate in X seconds ahead, and outputs a prediction of acceleration in X seconds ahead.

After the control device 100 outputs the acceleration prediction, a description of how another control device or device that receives the acceleration prediction changes the vehicle control will be omitted, but for example, it can be determined that the vehicle will not accelerate until X seconds ahead. In this case, it is possible to connect to the ECM (Engine Control Module) 109 in FIG.

FIG. 2 illustrates a model for predicting driver acceleration/deceleration by applying the present invention to the control device 100 shown in FIG. Here, as a definition of the driver's acceleration/deceleration prediction, it is defined as acceleration when the driver operates the accelerator three seconds after the current time and the value becomes equal to or greater than the accelerator opening just before. In addition, 3 seconds after the current time, it is defined as not accelerating if the driver maintains the accelerator position and maintains the same degree of accelerator opening as immediately before.

For example, two model examples are shown as methods of predicting whether the driver will accelerate after 3 seconds.

The first prediction model 1 is a model that inputs the vehicle speed and inter-vehicle distance, compares them with the target, and calculates the driver's required acceleration several seconds ahead. The calculation formula for this model is given by the following formula.

α: maximum acceleration Vf: own vehicle speed Vdes: target speed ΔX: inter-vehicle distance Xsigma: target inter-vehicle distance

In formula (1), the control device 100 determines that the vehicle should be accelerated if the vehicle-to-vehicle distance is greater than the target vehicle-to-vehicle distance, and that the vehicle should not be accelerated if it is shorter than the target vehicle-to-vehicle distance.

For example, when the vehicle speed of the own vehicle is the same as the target vehicle speed, that is, when Vf/Vdes = 1, and the vehicle-to-vehicle distance is far from the target, the term related to the vehicle-to-vehicle sensitivity, Xsigma/ΔX, works positively, Acceleration is positive.

When the inter-vehicle distance matches the target, that is, when Xsigma/ΔX=1, and the vehicle speed exceeds the target vehicle speed, the term Vf/Vdes, which is related to vehicle speed sensitivity, becomes negative. Working acceleration becomes negative.

The control device 100 predicts whether the driver will accelerate after 3 seconds depending on whether the acceleration is positive or negative.

The second prediction model 2 is a model that runs so that the vehicle speed of the preceding vehicle is within a predetermined value relative to the vehicle speed of the own vehicle. The calculation formula for this model is given by the following formula.

Vf: own vehicle speed Vp: preceding vehicle speed thsc: speed maintenance determination threshold thv: acceleration/deceleration determination threshold

First, in equations (2) and (3), it is confirmed that the difference between the vehicle speed Vp of the preceding vehicle and the vehicle speed Vf of the own vehicle is equal to or less than the speed maintenance determination threshold, and the vehicle is driven within a certain relative speed range with respect to the preceding vehicle. The control device 100 determines whether the The logic is to predict that the driver will not accelerate if the vehicle is driving within a certain range of relative speed to the preceding vehicle.

In formula (2), the control device 100 predicts that the vehicle will accelerate when the vehicle speed of the own vehicle is lower than the vehicle speed Vp of the preceding vehicle and the vehicle speed of the preceding vehicle is greater than or equal to the acceleration/deceleration determination threshold thv.

Further, in the expression (3), when the vehicle speed of the own vehicle is higher than the vehicle speed Vp of the preceding vehicle and the vehicle speed of the preceding vehicle is lower than the acceleration/deceleration determination threshold thv, the vehicle is decelerated, that is, the vehicle must be accelerated. The controller 100 predicts.

Inputs from the sensors shown in FIG. 1 are given to these two prediction models, and control device 100 predicts whether or not the vehicle will accelerate after 3 seconds.

Fig. 3 shows a block diagram of the first embodiment of the present invention.

The two prediction models in Fig. 2 are incorporated as prediction model 1 (block 301) and prediction model 2 (block 302) in the block diagram (Fig. 3), respectively.

As described above, the inter-vehicle distance and the vehicle speed are input to prediction model 1 and prediction model 2. Based on this input, prediction model 1 and prediction model 2 each determine whether the driver will accelerate after 3 seconds, and output the prediction result.

On the other hand, the correct state generation unit 303 in the block diagram receives the accelerator opening degree at the present time, compares it with the accelerator opening degree immediately before, determines whether the accelerator is stepped on, that is, accelerates, and converts the correct state into the correct state. The generating unit 303 generates. When comparing the output of the correct state generation unit 303 and the prediction result output from each prediction model, the output of each prediction model is the prediction result after 3 seconds, so the judgment result of the correct state generation unit 303 is different. In order to synchronize the time, the prediction result of the prediction model is stored in the prediction result storage unit 304 so that the prediction result of 3 seconds before is output.

The prediction result correct answer determination unit 305 compares the prediction result output from the prediction result storage unit 304 three seconds ago with the correct state calculated by the correct state generation unit 303, and determines whether the prediction result of each prediction model is correct. determine whether

The prediction result correct answer determination unit 305 outputs this correct/incorrect answer information to the arbitration parameter storage unit 306 .

The arbitration parameter storage unit 306 stores the number of correct and incorrect answers determined by the prediction result correct answer determination unit 305 for each prediction model.

Each prediction model correct answer rate calculation unit 307 calculates the correct answer rate from the number of correct answers and the number of incorrect answers of each prediction model.

Each prediction model prediction experience value calculation unit 308 counts how many times each prediction model has made predictions and calculates a prediction experience value.

The prediction result arbitration unit 309 predicts the prediction results (301 to 302) of each prediction model, each prediction model correct answer rate calculated by each prediction model correct answer rate calculation unit 307, and each prediction model prediction empirical value calculation unit 308 calculated Using the prediction model prediction empirical value, the evaluation value of the prediction accuracy of each prediction model is calculated, and the prediction result of the prediction model with high prediction accuracy is calculated as the final prediction result.

A specific example of the processing of the prediction result correct answer determination unit 305 and the arbitration parameter storage unit 306 in FIG. 3 will be described using FIG.

In Figure 4, the confusion matrix represents the state of correct and incorrect answers.

For example, as shown in (a) of FIG. 4, when the current state is "accelerate", if the prediction result of each prediction model three seconds ago predicted "accelerate", it is correct. judge.

On the other hand, as shown in FIG. 4B, when the current state is "do not accelerate", if the prediction result of each prediction model three seconds ago predicted "accelerate", it is an incorrect answer. judge.

Also, as shown in (c) of FIG. 4, when the current state is "accelerate", if the prediction result of each prediction model three seconds ago is "not accelerated", it is an incorrect answer. judge.

Various methods of expression can be adopted as the method for calculating the percentage of correct answers. In order to calculate the correct answer rate of "accelerate", here, the predicted "accelerate" and the actual "accelerate" shown in Fig. 4-(a) are set as TP (True Positive), ), the predicted "accelerate" and the actual "not accelerated" are the first incorrect FP (False Positive), and the predicted "not accelerated" and the actual "accelerated" shown in Fig. 4-(c) are the second is set as an arbitration parameter used for calculating the percentage of correct answers, which will be described later.

There are various ways to define the correct state of TN, and since it is not necessary to explain the embodiment of the present invention, it is omitted here.

Also, regarding the prediction that the vehicle will accelerate after 3 seconds, the definitions of correct and incorrect answers when the vehicle accelerates after 3.1 seconds will not be part of the description of the embodiment of the present invention. For example, when the prediction of three seconds before is obtained from the prediction result storage unit 304 and the prediction result and the correct state are compared by the prediction result correct answer determination unit 305, part of the output of the prediction result storage unit 304 is processed for comparison. It is possible to solve the problem of extremely low TP by implementing this.

The prediction result correct answer determination unit 305 determines the states of TP, FN, and FP, and outputs the determined states to the arbitration parameter storage unit 306 . The arbitration parameter storage unit 306 stores the number of times of TP, FN, and FP determined by the prediction result correct answer determination unit 305 for each prediction model.

Each prediction model correct answer rate calculation unit 307 calculates the correct answer rate of each prediction model. Various calculation methods are conceivable for the percentage of correct answers, but here, the F value that can be calculated using the values of TP, FN, and FP described above is applied to the percentage of correct answers. The calculation method of the F value is given by the following formula.
Recall rate (Recall) = TP/(TP + FN)
Precision = TP/(TP + FP)
F value = (2 x Precision x Recall) / (Precision + Recall)

Here is an example of calculating the F value and applying it as the correct answer rate for the prediction of "accelerate" 3 seconds ahead. Considering the influence of downstream control, various calculation formulas can be considered as the percentage of correct answers, and an index other than the calculation method given as an example may be used as the percentage of correct answers.

The first feature of the embodiment of the present invention is that, in addition to the correct answer rate of each prediction model, the prediction empirical value is used for arbitration. The prediction empirical value stores the number of predictions in each prediction model, and expresses the reliability of the correct answer rate of the prediction result of the prediction model according to the number of predictions.

For example, when the correct answer rate is 100%, the correct answer is 1 out of 1 prediction, and the correct answer is 100 out of 100 predictions, the reliability of the correct answer rate is different.

Here, with regard to how to use the prediction empirical value, for example, when the prediction is performed 300 times or more, the prediction result of the prediction model can be trusted, for example, by using the definition that It is possible to prevent the accuracy rate of the prediction result of the acceleration/deceleration prediction control from decreasing before and after the control.

Although details will be described later, the microcomputer 101 (processor) of the present embodiment uses a plurality of prediction models at a first timing (current time) to predict a second timing (future model) from the state observed at that time. time) is stored in the RAM 102 (first memory), and a correct state indicating the correct state for the prediction result is generated from the state observed at the second timing (at the current time). do. The microcomputer 101 compares the correct state and the prediction result for each prediction model, calculates the correct answer rate for each prediction model, and calculates the prediction empirical value for each prediction model, which indicates the number of predictions made for each prediction model. do. The microcomputer 101 selects and outputs one prediction result from a plurality of prediction results using the correct answer rate for each prediction model and the prediction empirical value for each prediction model. As a result, highly reliable prediction results are output.

In order to express the number of predictions, the second feature of this embodiment is to calculate the number of predictions as the sum of the number of correct answers and the number of incorrect answers of each prediction model output by the prediction result correct answer determination unit 305. That is. That is, the prediction empirical value for each prediction model is the sum of the number of correct answers and the number of incorrect answers for each prediction model. This eliminates the need for a counter for counting the number of predictions.

An example of calculating the predicted empirical value from the sum of the number of correct answers and the number of incorrect answers will be explained using FIG. (a), (b), and (c) of FIG. 5 are the values of TP, FN, and FP of prediction model 1 and prediction model 2, respectively, stored in the arbitration parameter storage unit. Calculating the sum of these values results in (g) in FIG. Also, the tables shown in (I), (II), and (III) of FIG. 5 show how each prediction model repeats prediction and TP, FN, and FP are added.

In (I) of FIG. 5, the sum of TP, FN, and FP of prediction model 1 and prediction model 2 is 40. In (f) of (I) of FIG. 5, the F value of prediction model 1 is 0.75, and the F value of prediction model 2 is 0.77. On the other hand, in (II) of FIG. 5, the F value of the prediction model 1 is 0.89, and the F value of the prediction model 2 is 0.82. In addition, in (III) of FIG. 5, the F value of the prediction model 1 is 0.89, and the F value of the prediction model 2 is 0.82, and the correct answer rate of each prediction model converges by repeating the number of times.

As shown in the example of FIG. 5, when the correct answer rate (f) in the state (I) with little prediction experience is used to mediate the prediction results, the true correct answer rate of the prediction model 1 and the prediction model 2 is unknown. In some situations, it is possible to adopt the results of the erroneous predictive model.

Therefore, the sum of the arbitration parameters (TP, FN, FP) shown in (g) is calculated as a prediction empirical value, and when the prediction empirical value of each prediction model is a predetermined value or more, the prediction result of each prediction model is If configured to be reflected in the final prediction result, it is possible to prevent the prediction system from outputting an erroneous prediction. Further, in the case of the example of FIG. 5, it is possible to additionally acquire prediction empirical values without newly storing new calculation blocks or additional information.

Details will be described later, but the microcomputer 101 (processor) selects one prediction result by including the prediction result of the prediction model in the options when the prediction experience value of each prediction model exceeds the threshold, and each prediction If the prediction empirical value of the model is less than the threshold, one prediction result is selected without including the prediction results of the prediction model in the options. Here, the microcomputer 101 (processor) selects, for example, the prediction result of the prediction model with the highest percentage of correct answers.

In addition, the third feature of this embodiment is to store the number of times each prediction model made predictions, that is, the number of predictions made by each prediction model as a prediction empirical value. In other words, the prediction empirical value for each prediction model is the number of predictions made and counted for each prediction model. Thereby, for example, it is possible to easily confirm the temporal change of the prediction empirical value.

For example, if the conditions for permitting prediction are considered in advance in order to increase the prediction accuracy of each prediction model, each prediction model outputs prediction results only when the conditions for permitting prediction for each model are met. is used as the prediction empirical value. It is possible to confirm how much prediction is performed while each prediction model is allowed to make predictions, and whether the correct answer rate of the prediction model is reliable.

The prediction empirical value of each model obtained in this way is referred to by the prediction result arbitration unit 309 in the block diagram (FIG. 3), and the prediction result arbitration unit 309 calculates the prediction result, the prediction correct answer rate, and the prediction experience value of each prediction model. Based on this, it is decided which model's prediction result is to be used. In this way, in addition to considering the correct answer rate, by adding the predicted empirical value as a numerical value indicating the reliability of the correct answer rate, the value of the prediction model is not adopted even if the correct answer rate is high.

With the above measures, when starting up the system from the initial state, it is possible to prevent the correction of prediction results from lowering the correct answer rate of the entire prediction result due to the imperfect calculation result of the correct answer rate of each prediction model. is possible.

As described above, according to this embodiment, highly accurate future situation prediction is possible. In addition, it is possible to train multiple prediction models and select the one that matches the characteristics of the driver. For example, although the target speed and target inter-vehicle distance in FIG. 2 differ depending on the driver, general values are set as initial values. Here, the target speed and the target inter-vehicle distance are fitted to the characteristics of the driver (learning) through the process of validating the prediction result (Fig. 3, 304-307).

Also, it is possible to achieve both _CO2 reduction and drivability. Specifically, for example, when it is predicted that the vehicle will not accelerate, CO ₂ can be reduced by stopping the engine (coast stop). On the other hand, when it is predicted that the vehicle will accelerate, the acceleration responsiveness is improved and the drivability is improved by not stopping the engine.

In addition, we can provide a suitable platform for adding newly developed models to the vehicle. Specifically, for example, when a new prediction model is added as an add-on during the experiment phase or online update, the results of the new prediction model are excluded from the options until a certain amount of prediction experience is accumulated, and the conventional prediction Select the one with the highest percentage of correct answers from the model results.

[Second embodiment]
As a second embodiment, an example in which a third different prediction model 3 is added to the first embodiment is shown. FIG. 6 shows a block diagram of a state in which a prediction model 3 (block 310) is added in addition to the

prediction models

1 and 2 shown in the first embodiment.

In the example of prediction model 3, the acceleration of the vehicle is obtained from the acceleration sensor 107 mounted on the vehicle in FIG. 1, and signs of the acceleration turning from negative to positive are detected to predict "accelerate".

With the prediction model 1 and the prediction model 2 described in the first embodiment connected to the prediction system, the arbitration parameter of the prediction model 1 and the prediction model 2 has already been run several times and the past prediction situation and the information of the correct answer rate is accumulated.

On the other hand, when the prediction of the prediction model 3 is added, it is necessary to rewrite the program (ROM 103) of the control device 100 at least a little. The arbitration parameter, that is, the predicted correct answer rate, may be compromised.

A fourth feature of this embodiment is that the arbitration parameters are written to the non-volatile memory. That is, the microcomputer 101 (processor) stores the number of correct answers, the number of incorrect answers, and the prediction empirical value for each prediction model in the nonvolatile memory 104 (second memory). By storing the arbitration parameters of the prediction model 1 and the prediction model 2 in the non-volatile memory 104 of FIG. 1, the arbitration parameters of the prediction model 1 and the prediction model 2 are held even when the prediction model 3 is added.

In addition, by simultaneously holding the prediction empirical value, which is one of the arbitration parameters, in the nonvolatile memory, it is possible to prevent the initialization of the prediction results of the prediction model 1 and the prediction model 2, and the prediction accuracy in the conventional prediction can be improved. is maintained, the evaluation of prediction model 3 proceeds in the background, and when the prediction empirical value of prediction model 3 exceeds a predetermined value, it is reflected in the output of the final prediction result in the same way as

prediction models

1 and 2. configured as follows.

Fig. 7 shows how the correct answer rate and prediction empirical value of each prediction model change depending on the number of predictions when prediction model 3 is added.

(I) of FIG. 7 shows the state before the system starts prediction with prediction model 3 added. Arbitration parameters (a) to (c) of prediction model 1 and prediction model 2 are stored in a non-volatile memory and read out. In this state, the prediction of prediction model 3 has not yet been performed, and the prediction empirical value (g) of prediction model 3 is also zero.

(II) in FIG. 7 is the state after executing prediction 200 times from the state of (I). At this time, the prediction model 3 has the highest rate of correct answers (F value). For example, if the prediction experience of the prediction model is not 300 times or more, it is not reflected in the final prediction result. .

(III) in FIG. 7 is the state after performing 200 predictions from the state of (II). At this time, the correct answer rate of prediction model 3 is 0.79, which is lower than

prediction models

1 and 2. Then, although the percentage of correct answers was high at the time of (II), when the prediction was repeated many times, it turned out that the prediction accuracy of the prediction model 3 was low overall.

By holding the prediction empirical value in addition to the prediction correct answer rate in this way, even if the prediction model 3 outputs an incorrect prediction result in a state where the prediction model 3 has a high correct answer rate but the prediction reliability is low It is also possible to prevent the deterioration of the prediction accuracy of the whole system.

FIG. 8 is a flow chart diagram showing the flow of predictive control by the control device 100 in the first and second embodiments of the present invention. Hereinafter, each step will be described with reference to the block diagram of FIG. 6 and the flowchart of FIG.

(Fig. 8: Step S801)
The control device 100 starts predictive control.

(Fig. 8: Step S802)
Prediction model 1 (301), prediction model 2 (302), and prediction model 3 (310) of the prediction system each output a prediction result.

(Fig. 8: Step S803)
Each prediction model stores the prediction result in the prediction result storage unit 304 .

(Fig. 8: Step S804)
The correct state generation unit 303 generates a correct state.

(Fig. 8: Step S805)
The prediction result correct answer determination unit 305 refers to the prediction result of each past prediction model stored in the prediction result storage unit 304 in step 803 and compares it with the correct state generated in step 804 .

(Fig. 8: Step S806)
When the prediction result of the prediction model and the correct answer are compared in step 805 , if the prediction model is correct, the correct counter value of the prediction model [i] in the arbitration parameter storage unit 306 is incremented and stored in the arbitration parameter storage unit 306 . .

(Fig. 8: Step S807)
When the prediction result of the prediction model and the correct answer are compared in step 805, if the prediction model is incorrect, the incorrect counter value of the prediction model [i] in the arbitration parameter storage unit 306 is incremented and stored in the arbitration parameter storage unit. Store.

(Fig. 8: Step S808)
Each prediction model correct answer rate calculation unit 307 calculates the predicted correct answer rate of the prediction model [i] using the number of correct answers and the number of incorrect answers of each prediction model updated in step S806 or step S807.

(Fig. 8: Step S809)
The prediction empirical value is calculated by obtaining the sum of the arbitration parameters of each prediction model calculated in steps S806 and S807.

(Fig. 8: Step S810)
It is determined whether or not the prediction empirical value of each prediction model calculated in step S810 exceeds a predetermined threshold.

(Fig. 8: Step S811)
If the prediction empirical value is equal to or greater than a predetermined threshold in step S811, the prediction result of the prediction model [i] can be selected during arbitration.

(Fig. 8: Step S812)
If the prediction empirical value is less than the predetermined threshold in step S812, the prediction result of the prediction model [i] cannot be selected during arbitration.

(Fig. 8: Step S813)
In order to calculate the prediction correct answer rate and the prediction experience value of all the prediction models, it is determined whether the evaluation of the prediction correct answer rate and the prediction experience value using the arbitration parameter of all the prediction models has been completed.

(Fig. 8: Step S814)
In step S813, if calculation of the prediction correct answer rate and prediction experience value of all prediction models has not been completed, i is incremented and the prediction correct answer rate and prediction experience value of the next prediction model are calculated.

(Fig. 8: Step S815)
When the evaluation of the prediction correct answer rate and the prediction empirical value of all the prediction models is completed in step S813, the process proceeds to step S815.

In step S811, the product of the prediction result of the prediction model determined to be selectable at the time of arbitration based on the prediction empirical value of the prediction model and the prediction correct answer rate is calculated, and the prediction result of the highest prediction model is selected as the final prediction result of the prediction system.

The calculations related to the prediction system from step S801 to step S815 are included in the control program stored in the ROM 103 of the control device 100 and are repeatedly executed.

[Third Embodiment]
FIG. 9 is a block diagram for explaining the third embodiment to which the present invention is applied.

The fifth feature of the present embodiment is that the system of the first embodiment is further developed, a state determination unit 901 is added, and the prediction result of each prediction model and the correct answer are calculated by comparing the correct answer of each prediction model, The number of incorrect answers is stored in the state-based arbitration parameter storage unit 902 as the arbitration parameter of the prediction apparatus for each state determined by the state determination unit 901 .

A description will be given using an example of a system that predicts whether or not the vehicle will accelerate after 3 seconds, which was used to describe the first and second embodiments.

Prediction model 1 (block 301), prediction model 2 (block 302), prediction model (block 303), correct state generation unit 303, prediction result storage unit 304, prediction result correct answer determination unit 305 are the first embodiment and the second embodiment. It has the same function as the content explained in the form.

The state determination unit 901 determines the current state (situation). The current situation here may be an input used for the prediction model, an input for determining the correct state, or another input.

In the example of FIG. 9, the state determination unit 901 determines the current state from three inputs, and sends the state ID for each state to the arbitration parameter storage unit 902 for each state as a result of the state determination based on the determination.

In this example, the state determination unit generates a state ID based on information on vehicle speed, inter-vehicle distance, and relative speed. FIG. 10 shows how the vehicle speed, inter-vehicle distance, and relative speed are each discriminated into eight states and state IDs are generated.

For example, if the current vehicle speed is 50 km/h, the inter-vehicle distance is 40 m, and the relative speed is +0.5 km/h, the vehicle speed state is indicated as state 4, the inter-vehicle distance state is indicated as state 3, and the relative speed state is indicated as state 5. and these states are referred to as a bundling state ID 435 .

The state-by-state arbitration parameter storage unit 902 compares the prediction result of each prediction model with the correct answer to determine whether each prediction model is correct or incorrect, and then stores the arbitration parameter by state ID. For state ID 435, the arbitration parameter corresponding to 435 is updated.

In the system having the fifth feature of the present embodiment, calculation of each prediction model correct answer rate is performed by each prediction model correct answer rate calculation unit 903 using the arbitration parameters stored in the state-specific arbitration parameter storage unit 902 in FIG. Each prediction model prediction experience value calculation unit 904 calculates the prediction experience value of each prediction model.

At this time, the prediction result arbitration unit 905 arbitrates the prediction results using the prediction result of each prediction model, the prediction correct answer rate for each situation, and the prediction empirical value for each situation.

In other words, the microcomputer 101 (processor) stores a state ID which is composed of a set of IDs corresponding to the range to which each of the measured values from a plurality of sensors belongs and which indicates the state at the first timing (current time). calculate. The microcomputer 101 calculates the percentage of correct answers for each prediction model and each state ID, and calculates the prediction empirical value for each prediction model and each state ID. The microcomputer 101 outputs one prediction result from the plurality of prediction results using the correct answer rate for each prediction model and for each state ID and the prediction empirical value for each prediction model and each state ID.

According to the fifth feature of the present embodiment, for example, in a certain state, the prediction is good and the correct answer rate is higher than the other prediction model, but in another state the prediction is wrong, and the other prediction model When the percentage of correct answers is lower than , it is possible to prevent the percentage of correct answers from being averaged by the number of correct answers and incorrect answers in all states. As a result, if there is a prediction model that is superior to other prediction models in a certain state, even if the correct answer rate is low in other states, the prediction results will be reflected in the final output of the system in the area where that prediction model is good. It becomes possible to

A sixth feature of the present embodiment is to calculate a prediction empirical value using the number of correct and incorrect answers in predictions made so far by each prediction model stored in the state-specific arbitration parameter storage unit 902. .

According to the sixth feature of the present embodiment, it is possible to calculate the prediction empirical value without additionally storing information on the number of predictions, and the reliability of the correct answer rate of each prediction model calculated for each state can be confirmed, and the final output of the system can be determined by the prediction result of each prediction model, the correct answer rate of each prediction model by state, and the prediction empirical value of each prediction model by state.

A specific example will be explained using FIG. In the acceleration/deceleration predictive control described in the above embodiment, for example, in a region where the vehicle speed is low, it is assumed that the driver unconsciously drives with the preceding vehicle in mind.

(IV) in FIG. 11 shows the overall prediction accuracy rate of prediction models 1 to 3, where prediction model 2 is inferior to

prediction models

1 and 3. On the other hand, in the region of the state ID 243 selected in the states of vehicle speed state 2, inter-vehicle distance state 4, and relative speed state 3 ((V) in FIG. 11), the prediction model 2 is replaced by another prediction model. The correct answer rate is high.

In this state, the accuracy rate of the prediction result of prediction model 2 is high, and the prediction accuracy of the entire system can be expected to be improved by using the prediction result of prediction model 2 rather than using other prediction models.

Also, in areas where the vehicle speed is high, it is assumed that the driver drives with the intention of achieving the target speed rather than the awareness of the preceding vehicle. In state 642 of (VI) in FIG. 11, the result of prediction model 1 has the highest percentage of correct answers, and the prediction result of prediction model 1 is selected as the final prediction result.

　I want to output the final prediction result using the correct answer rate by state, but I want to check whether the correct answer rate is a reliable value. According to the sixth feature of the present embodiment, it is possible to check the number of predictions made by each prediction model in the state determined by the state determination unit. In this case, the prediction result of prediction model 1 can be output to the final output. In addition, even if the correct answer rate of prediction model 1 is high, if the prediction experience value is low, that is, if the prediction performance in that state is poor, it is possible not to reflect the result of prediction model 1 in the final result. be.

The seventh feature of this embodiment is to store the number of predictions made by the prediction model in that state as the prediction empirical value instead of the arbitration parameter of each prediction model, that is, the number of correct answers and the number of incorrect answers. .

For example, if the conditions for permitting prediction are considered in advance in order to increase the prediction accuracy of each prediction model, each prediction model outputs prediction results only when the conditions for permitting prediction for each model are met. is used as the prediction empirical value. Depending on the state, some prediction models may be allowed to perform predictions, and some prediction models may not be allowed to perform predictions. At that time, the arbitration parameters of the disapproved prediction model are not updated.

In such a configuration, the arbitration parameter can prevent the prediction result from being incorrect and the correct answer rate from decreasing when the prediction range assumed in advance by each prediction model is exceeded. The prediction accuracy of the entire system can be expected to be improved by segregating the prediction regions intended by the prediction model.

FIG. 12 is a flow chart diagram showing the flow of predictive control by the control device 100 according to the third embodiment of the present invention. Hereinafter, each step will be described with reference to the block diagram of FIG. 9 and the flow chart of FIG.

(Fig. 12: Step S1201)
The control device 100 starts predictive control.

(Fig. 12: Step S1202)
Prediction model 1 (301), prediction model 2 (302), and prediction model 3 (310) of the prediction system each output a prediction result.

(Fig. 12: Step S1203)
Each prediction model stores the prediction result in the prediction result storage unit 304 .

(Fig. 12: Step S1204)
The correct state generation unit 303 generates a correct state.

(Fig. 12: Step S1205)
A state determination unit 901 determines the current state and selects a state ID.

(Fig. 12: Step S1206)
The prediction result correct answer determination unit 305 refers to the prediction result of each past prediction model stored in the prediction result storage unit 304 in step 1203 and compares it with the correct state generated in step 1204 .

(Fig. 12: Step S1207)
When the prediction result of the prediction model and the correct answer are compared in step 1206, if the prediction model is correct, the correct counter value of the prediction model [i] for the state ID selected in step S1205 is stored in the state-by-state arbitration parameter storage unit 902. is incremented and stored in the state-specific arbitration parameter storage unit 902 .

(Fig. 12: Step S1208)
When the prediction result of the prediction model and the correct answer are compared in step 1206, if the prediction model is incorrect, the incorrect answer counter of the prediction model [i] in the state ID selected in step S1205 of the state-by-state arbitration parameter storage unit 902 The value is incremented and stored in the arbitration parameter storage unit.

(Fig. 12: Step S1209)
Each prediction model correct answer rate calculation unit 903 calculates the predicted correct answer rate of the prediction model [i] using the number of correct answers and the number of incorrect answers of each prediction model for each state updated in step S1207 or step S1208.

(Fig. 12: Step S1210)
Each prediction model prediction empirical value calculation unit 904 calculates the prediction empirical value by obtaining the sum of the arbitration parameters of each prediction model for each state calculated in steps S1206 and S1207.

(Fig. 12: Step S1211)
It is determined whether or not the prediction empirical value of each prediction model calculated in step S1210 exceeds a predetermined threshold.

(Fig. 12: Step S1212)
If the prediction empirical value is equal to or greater than the predetermined threshold in step S1211, the prediction result of the prediction model [i] can be selected during arbitration. That is, when the prediction empirical value of each prediction model exceeds the threshold, the microcomputer 101 (processor) selects one prediction result including the prediction result of that prediction model in the options. This ensures that prediction results with low reliability are not selected.

(Fig. 12: Step S1213)
If the prediction empirical value is less than the predetermined threshold in step S1212, the prediction result of the prediction model [i] cannot be selected during arbitration. That is, when the prediction empirical value of each prediction model is less than the threshold, the microcomputer 101 (processor) selects one prediction result without including the prediction result of that prediction model in the options. Thereby, a highly reliable prediction result is selected.

(Fig. 12: Step S1214)
In order to calculate the prediction accuracy rate and the prediction experience value of all prediction models, it is determined whether the evaluation of the prediction accuracy rate and the prediction experience value using the state-specific arbitration parameters of all prediction models has been completed.

(Fig. 12: Step S1215)
In step S1214, if calculation of the prediction correct answer rate and prediction experience value of all prediction models has not been completed, i is incremented and the prediction correct answer rate and prediction experience value of the next prediction model are calculated.

(Fig. 12: Step S1216)
When the evaluation of the prediction accuracy rate and the prediction empirical value of all the prediction models is completed in step S1213, the process moves to step S1215.

In step S1212, the product of the prediction result of the prediction model determined to be selectable at the time of arbitration based on the prediction empirical value of the prediction model and the prediction correct answer rate is calculated, and the prediction result of the highest prediction model is selected as the final prediction result of the prediction system.

The calculations related to the prediction system from step S1201 to step S1216 are included in the control program stored in the ROM 103 of the control device 100 and are repeatedly executed.

In the above embodiment, in a system that predicts an event using a plurality of prediction models, the prediction correct answer rate is calculated for the prediction result calculated by each prediction model for the purpose of improving the prediction accuracy of the entire system, In addition, the calculation of the prediction empirical value for confirming whether the prediction correct answer rate is a reliable value was explained.

Here, the threshold for evaluating the prediction empirical value in S810 of FIG. 8 or S1211 of FIG. good too. That is, the threshold may be a common value for all prediction models, or the threshold may be set for each prediction model.

For example, in the second or third embodiment, the prediction model 1 and the prediction model 2 are evaluated in advance, configured to be provided as the minimum prediction function of the system, and it is desired to start prediction from the time of system startup. is also conceivable. In such a case, any prediction result can be adopted as a candidate when the final prediction result is selected by setting the initial value of the prediction empirical value in advance to be equal to or greater than the number of selections that can be made at the time of arbitration. In other words, the number of correct answers, the number of incorrect answers, or the initial value of the prediction empirical value for each prediction model is any number other than zero.

With the prediction empirical value of the above embodiment, a threshold is used to check whether or not prediction has been executed a predetermined number of times, and if the threshold is not exceeded, selection is disabled during arbitration. If the prediction result of each prediction model is to be included as an option for the final prediction result regardless of the magnitude of the prediction empirical value, the value used for the threshold should be set to 0 or less. Specifically, for example, the initial value of the number of correct answers, the number of incorrect answers, or the prediction empirical value for each prediction model is zero.

By using this threshold value as a parameter for calculating the prediction empirical value and having a different value for each prediction model, for example, in the example of the second embodiment, the predictions of the prediction model 1 and the prediction model 2 are predicted in the prediction result arbitration unit from the first time Used to select the result, if the parameter for calculating the prediction empirical value of the prediction model 3 is set to an arbitrary number of times, only the prediction model 3 will be included in the prediction result options after the prediction results are accumulated. It is possible.

In addition, in the prediction empirical value of the above embodiment, a threshold is used to check whether or not prediction has been executed a predetermined number of times, and if the threshold is not exceeded, the prediction result cannot be selected during arbitration. Until the determination of the threshold value is performed in S809 of FIG. 8 or S1210 of FIG. Similar arbitration can be performed by modifying the flow chart so as to calculate the product of the prediction result, the prediction correct answer rate, and the prediction empirical value in step S815 or S1216 of prediction result arbitration.

In other words, the microcomputer 101 (processor) (i) selects the prediction result of the prediction model with the highest product of the prediction result value and the correct answer rate, or (ii) compares the prediction result value, the correct answer rate, and the prediction The prediction result of the prediction model with the highest product with the empirical value may be selected. Thereby, a highly reliable prediction result is selected.

In addition, as another prediction result arbitration method, a method of performing calculation by setting the number of prediction executions in the numerator and setting the parameter for calculating the prediction experience value in the denominator, and calculating the prediction experience value in decimals is also conceivable. . Even in this case, it is possible to change the degree of contribution of the prediction result of each prediction model to the selection of the final result by setting the prediction empirical value calculation parameter for each prediction model in advance.

For example, in the acceleration/deceleration predictive control described in the above embodiment, when predicting acceleration after 3 seconds, the prediction result of prediction model 1 is "accelerate" correct answer rate 0.8, prediction experience value 1, prediction When the prediction result of model 2 is "not accelerated", the correct answer rate is 0.6, the prediction experience value is 1, and the prediction result of prediction model 3 is "accelerate", the correct answer rate is 0.9, and the prediction experience value is 0.5 , the product of the correct answer rate and the predicted experience value of the prediction model 1 is 0.8, the product of the correct answer rate and the prediction experience value of the prediction model 2 is 0.6, and the product of the correct answer rate and the prediction experience value of the prediction model 3 is 0.8. 45, and although the correct answer rate of the prediction model 3 is high, the prediction result of the prediction model 1 is supported because the prediction experience is small.

If the correct answer rate of prediction model 3 is 0.9 and the prediction experience value is 0.9, the product of the correct answer rate and the prediction experience value is 0.81, and the situation of prediction model 1 and prediction model 2 described above When compared, the evaluation value is the highest, and the result of prediction model 3 is supported.

In other words, the microcomputer 101 (processor) selects the prediction result of the prediction model with the highest product of the correct answer rate and the prediction empirical value. Thereby, a highly reliable prediction result is selected.

In this way, when the prediction empirical value is expressed in decimals, even if the prediction performance is low, it is possible to configure the prediction result to be reflected in the final output when the correct answer rate is high.

Although the embodiments of the present invention have been described in detail above, the present invention is not limited to the above-described embodiments, and various designs can be made without departing from the spirit of the invention described in the claims. Changes can be made. For example, the above-described embodiments have been described in detail in order to explain the present invention in an easy-to-understand manner, and are not necessarily limited to those having all the described configurations.

For example, the control that predicts whether the vehicle will accelerate after 3 seconds, which is mentioned in the embodiment, can be applied to predict the future state from the current state. It can also be applied to /G (Wastegate) control and the like. In addition, regarding the control for predicting whether to accelerate after 3 seconds as described in the embodiment, the plurality of models may be composed of neural network models and models based on other statistical theories in addition to the prediction models described in the embodiment. good.

In the above embodiment, the camera sensor 105 measures the inter-vehicle distance, but other sensors such as millimeter wave radar may be used to measure the inter-vehicle distance.

The state indicated by the state ID in the third embodiment is three-dimensional (vehicle speed, inter-vehicle distance, relative speed), but may be one-dimensional, two-dimensional, or four or more dimensions. In addition to the vehicle speed, inter-vehicle distance, and relative speed, the vehicle speed of the preceding vehicle, acceleration of the vehicle, accelerator opening, altitude, and the like may be used.

In addition, although block diagrams and flowcharts are used to explain the present invention, each function and each step of the drawings used for explanation may be realized by software or hardware. Some may be implemented in either.

In addition, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with another configuration.

It should be noted that the embodiment of the present invention may have the following aspects.

(1). A plurality of prediction models that predict the future state from the current state and output prediction results, a prediction result storage unit for storing each prediction result predicted by the plurality of prediction models in the past, and a correct state from the current state and a prediction result correct determination unit that compares the correct state and past prediction results of a plurality of prediction models stored in the prediction result storage unit and determines whether or not the prediction of each prediction model is correct. a prediction model arbitration parameter storage unit for storing the number of correct answers and the number of incorrect answers of each prediction model determined by the correct prediction result judgment unit; A prediction model correct answer rate calculation unit that calculates the correct answer rate of each prediction model using the number of correct answers, and a prediction result arbitration unit that outputs one prediction result using the prediction result of each prediction model and the correct answer rate of each prediction model The system further comprises a prediction empirical value calculation unit that counts how many times each prediction model has made predictions and calculates the prediction experience value of each prediction model, and the prediction result arbitration unit is , a system characterized by outputting one prediction result using the prediction result of each prediction model, the correct answer rate of each prediction model, and the prediction empirical value of each prediction model.

(2). In the system of (1), the prediction experience value calculation unit calculates the prediction experience value based on the sum of the number of correct answers and the number of incorrect answers of each prediction model stored in the prediction model mediation parameter storage unit. system.

(3). In the system of (1), the prediction empirical value calculation unit calculates the prediction empirical value by counting the number of times each prediction model executes the process.

(4). In the system of (1), the prediction model arbitration parameter storage unit writes the arbitration parameter of each prediction model in a non-volatile memory.

(5). In the system of (1), the system further includes a state determination unit that determines the current state and calculates a state ID corresponding to the current state, and the prediction model determined by the prediction result correct determination unit When the prediction model mediation parameter storage unit for each state is provided for storing the number of correct answers and the number of incorrect answers for each state selected by the state ID, and the prediction result mediation unit mediates the prediction results calculated by each prediction model. , the prediction correct answer rate by state calculated using the arbitration parameter of each prediction model stored in the prediction model arbitration parameter storage unit by state, and the prediction model arbitration parameter storage unit by state Calculate the prediction empirical value by state calculated using the arbitration parameter, and use the prediction result of each prediction model, the prediction correct answer rate by state of each prediction model, and the prediction empirical value by state of each prediction model, A system characterized by outputting one prediction result.

(6). In the system of (5), the prediction experience value calculation function is characterized in that the prediction experience value is calculated based on the sum of the number of correct answers and the number of incorrect answers for each prediction model, which are stored in the prediction model mediation parameter storage function. system.

(7). In the system of (5), the prediction empirical value calculation function counts the number of times each prediction model executes the process and calculates the prediction empirical value.

(8). In the system of (5), the prediction model arbitration parameter storage function writes the arbitration parameter of each prediction model into a non-volatile memory.

(9). In the system of (4), the initial value of the arbitration parameter of each prediction model stored in the non-volatile memory is set in advance to any number other than 0.

(10). In the system of (4), the initial value of the arbitration parameter of each prediction model stored in the nonvolatile memory is set to 0 in advance.

(11). In the system of (8), the initial value of the state-specific arbitration parameter stored in the nonvolatile memory is set in advance to any number other than 0.

(12). In the system of (8), the initial value of the state-specific arbitration parameter stored in the non-volatile memory is set to 0 in advance.

(13). In the system of (2) or (3), when the prediction experience value for each prediction model calculated by the prediction experience value calculation function is compared with an arbitrary threshold, and the prediction experience value of each prediction model exceeds an arbitrary threshold (2) A system characterized in that the prediction result of each prediction model arbitrated by the prediction result arbitration unit is included in the arbitration options.

(14). In the system of (13), a system characterized by using a common value for each prediction model as an arbitrary threshold to be compared with the prediction empirical value of each prediction model.

(15). In the system of (13), a system characterized by using different values for each prediction model as arbitrary thresholds to be compared with prediction empirical values of each prediction model.

(16). In the system of (6) or (7), the prediction experience value by state of each prediction model calculated by the prediction experience value calculation function is compared with an arbitrary threshold, and the prediction experience value by state of each prediction model is an arbitrary A system, wherein the prediction result of each prediction model arbitrated by the prediction result arbitration unit is included in the arbitration options when the threshold value is exceeded.

(17). In the system of (16), a system characterized by using a common value for each prediction model as an arbitrary threshold to be compared with the state-specific prediction empirical value of each prediction model.

(18). In the system of (16), a system characterized by using a different value for each prediction model as an arbitrary threshold to be compared with the state-specific prediction empirical value of each prediction model.

In the system of the above embodiment configured as in (1) to (18), a plurality of prediction models estimate the future state and output the prediction results, and the correct answer rate of the prediction results of each prediction model and take into account predictive experience to determine whether the percentage of correct answers is reliable. Therefore, when an existing system is expanded and a prediction model based on a new concept is added, it is possible to prevent a decrease in prediction accuracy due to the side effect of adding the prediction model.

In addition, by subdividing and storing the arbitration parameters of each prediction model for each situation, it is possible to improve the prediction accuracy of the entire system by appropriately switching the prediction model for each situation.

DESCRIPTION OF SYMBOLS 100... Control apparatus 101... Microcomputer 104... Non-volatile memory 105... Camera sensor 106... Vehicle speed sensor 107... Acceleration sensor 108... Accelerator opening sensor 109... ECM
110...

Engine

301, 302...Prediction model 303...Correct state generation unit 304...Prediction result storage unit 305...Prediction result correct answer determination unit 306...Arbitration parameter storage unit 307...Prediction model correct answer rate calculation unit 308...Prediction model prediction empirical value calculation Unit 309 Prediction result arbitration unit 310 Prediction model 901 State determination unit 902 State arbitration parameter storage unit 903 Prediction model correct answer rate calculation unit 904 Prediction model prediction empirical value calculation unit 905 Prediction result arbitration unit

Claims

A prediction device comprising a processor and a first memory,
The processor
storing in the first memory a prediction result indicating a state at a second timing predicted from a state observed at that time using each of the plurality of prediction models at the first timing;
generating a correct state indicating a correct state for the prediction result from the state observed at the second timing;
Comparing the correct state and the prediction result for each prediction model, calculating a correct answer rate for each prediction model,
calculating a prediction empirical value for each prediction model that indicates the number of predictions performed for each prediction model;
A prediction device that selects and outputs one prediction result from a plurality of prediction results using the correct answer rate for each prediction model and the prediction empirical value for each prediction model.
A prediction device according to claim 1,
The prediction empirical value for each prediction model is
A prediction device characterized by being the sum of the number of correct answers and the number of incorrect answers for each of the prediction models.
A prediction device according to claim 1,
The prediction empirical value for each prediction model is
It is the number of predictions performed and counted for each of the prediction models.
A prediction device according to claim 1,
comprising a non-volatile second memory;
The processor
A prediction device, wherein the number of correct answers, the number of incorrect answers, and the prediction empirical value for each of the prediction models are stored in the nonvolatile second memory.
A prediction device according to claim 1,
The processor
calculating a state ID composed of a set of IDs corresponding to the range to which each of the measured values from a plurality of sensors belongs and indicating the state at the first timing;
calculating the percentage of correct answers for each of the prediction models and each of the state IDs;
calculating a prediction empirical value for each of the prediction models and each of the state IDs;
One prediction result is output from the plurality of prediction results using the correct answer rate for each prediction model and each state ID and the prediction empirical value for each prediction model and each state ID. prediction device.
A prediction device according to claim 1,
The prediction device, wherein the number of correct answers, the number of incorrect answers, or the initial value of the prediction empirical value for each of the prediction models is an arbitrary number other than zero.
A prediction device according to claim 1,
The prediction device, wherein an initial value of the number of correct answers, the number of incorrect answers, or the prediction empirical value for each of the prediction models is zero.
A prediction device according to claim 1,
The processor
When the prediction empirical value of each of the prediction models is equal to or greater than a threshold, including the prediction results of the prediction model in the options and selecting one prediction result,
If the prediction empirical value of each of the prediction models is less than the threshold, selecting one prediction result without including the prediction results of the prediction model in the options;
A prediction device characterized by:
A prediction device according to claim 8,
The threshold is
A prediction device characterized by being a common value for all the prediction models.
A prediction device according to claim 8,
The threshold is
A prediction device characterized by being set for each of the prediction models.
A prediction device according to claim 8,
The processor
(i) Selecting the prediction result of the prediction model with the highest correct answer rate, or (ii) Selecting the prediction result of the prediction model with the highest product of the correct answer rate and the prediction empirical value. prediction device.
A prediction device according to claim 8,
The processor
(i) Selecting the prediction model with the highest product of the prediction result value and the correct answer rate, or (ii) Selecting the prediction model with the highest product of the prediction result value, the correct answer rate, and the prediction experience value. A prediction device characterized by:
A prediction method executed by the prediction device according to claim 1,
a step of storing a prediction result indicating a state at a second timing predicted from a state observed at that time using each of the plurality of prediction models at the first timing;
generating a correct state indicating a correct state for the prediction result from the state observed at the second timing;
Comparing the correct state with the prediction result for each prediction model, and calculating a correct answer rate for each prediction model;
calculating a prediction empirical value for each prediction model that indicates the number of predictions made for each prediction model;
selecting and outputting one prediction result from a plurality of prediction results using the correct answer rate for each prediction model and the prediction empirical value for each prediction model;
Forecasting methods, including