TWM625736U - Ensemble learning predicting system - Google Patents

Abstract

An ensemble learning predicting system includes a base predictor training model, a predictor weighting function training module, an evaluation module, a sample weight adjustment module. The predictor weighting function training module is connected to the base predictor training model. The sample weight adjustment module is connected to the evaluation module and is connected to the predictor weighting function training module. In the system, the following operations are performed: based on a plurality of sample weights of a plurality of sample data and respective identification results on the sample data by a plurality of predictors, establishing respective confidence models of the predictors; based on the confidence models, calculating respective scores of the confidence models, and sorting the confidence models based on the scores to select one confidence model among the confidence models as a target confidence model at a current iteration round; based on the confidence models and the identification results, adjusting the sample weights of the sample data; and based on the sample data and the adjusted sample weights, predicting the confidence models at a next round until all iteration rounds are completed to obtain a plurality of target confidence models. The target confidence models are used for generating a prediction result of a test data.

Description

Holistic Learning Prediction System

This creation is about a holistic learning prediction system.

Predicting future events by analyzing historical data plays an important role in manufacturing, as well as other industries. In a smart manufacturing factory, the machine will record many production parameters, such as temperature, pressure, gas flow, etc. These production parameters can be used as sample data to predict product quality or whether the machine will fail.

Ensemble learning is one of the supervised learning methods in machine learning methods. In holistic learning, multiple predictors (or hypothesis hypothesis) are combined with individual weights, and the hypothesis with higher confidence is dynamically selected to obtain the prediction result.

Figure 1 shows an example of holistic learning. As shown in Fig. 1, the sample data x is input to the basic predictors h ₁ ( x ), h ₂ ( x ) and h ₃ ( x ) respectively, wherein the basic predictors h ₁ ( x ), h ₂ ( x ) The weights with h ₃ ( x ) are w ₁ , w ₂ and w ₃ , respectively. If the weights are fixed (that is, the weights do not vary with x), the prediction result is y = w ₁ h ₁ ( x ) + w ₂ h ₂ ( x ) + w ₃ h ₃ ( x ). And if the weights are dynamic (that is, the weights w _i = g _i ( x ) , and w _i varies with x ), then the prediction result y = g ₁ ( x ) h ₁ ( x ) + g ₂ ( x ) h ₂ ( x ) + g ₃ ( x ) h ₃ ( x ). In ensemble learning, dynamic weights usually lead to better prediction results. However, if the complexity of the sample distribution is high, the confidence estimation is likely to be inaccurate, and it is difficult to effectively train dynamic weights.

Therefore, how to effectively train dynamic weights even when the sample space is complex is one of the key points in the design of an integrated learning test system in the industry.

According to this creation, an integrated learning and prediction system is proposed, which includes a basic predictor training model, a predictor weight function training module, an evaluation module, and a sample weight adjustment module. The basic predictor training model uses a plurality of training data transmitted from a signal source to establish a plurality of basic predictors, wherein the signal source is a sensor. The predictor weight function training module is connected to the basic predictor training model, initializes multiple sample weights of multiple sample data, and initializes a processing set. After initialization, the processing set includes multiple predictors corresponding to these basic predictors , and, in a first iteration round, according to the sample data and the sample weights, a plurality of predictor weight functions for the predictors in the processing set are established, and each of these predictors in the processing set is The predictor respectively predicts the sample data to generate a prediction result for each of the sample data. The evaluation module evaluates the predictor weight functions, and according to an evaluation result, selects a target predictor weight function from the predictor weight functions established in the first iteration round and selects a corresponding target predictor from the processing set target predictor, one of the predictor weight functions. The sample weight adjustment module is connected to the evaluation module and the predictor weight function training module to update the processing set and update the sample weights of the sample data. The updated processing set does not include the target predictor. Wherein, according to the sample weights of the updated sample data, the next iteration is performed to repeat the above operation to select another target predictor from the updated processing set and obtain another target predictor weight function, until all iterative rounds are performed to select a plurality of target predictor weight functions and a plurality of target predictors to integrate into an overall predictor, wherein the overall predictor includes these target predictor weight functions and these target predictors, and a prediction of the overall predictor The result is further displayed on a display.

In order to have a better understanding of the above-mentioned and other aspects of the present creation, the following examples are given and described in detail with the accompanying drawings as follows:

210-280: Steps

300: Holistic Learning Test System

310: Basic Predictor Training Model

320: Predictor weight function training module

330: Evaluation Module

340: Sample weight adjustment module

Figure 1 shows an example of holistic learning.

FIG. 2 shows a flow chart of an integrated learning prediction method according to an embodiment of the present application.

FIG. 3 shows a functional block diagram of an integrated learning and testing system according to another embodiment of the present application.

The technical terms in this specification refer to the common terms in the technical field. If some terms are described or defined in this description, the interpretations of these terms are subject to the descriptions or definitions in this description. Each embodiment of the present disclosure has one or more technical features. Under the premise of possible implementation, those skilled in the art can selectively implement some or all of the technical features in any embodiment, or selectively combine some or all of the technical features in these embodiments.

FIG. 2 shows a flowchart of an integrated learning prediction method according to an embodiment of the present application, which can be applied to an electronic device with a processor. As shown in Fig. 2, in step 210 (training phase), a set of basic predictors h ₁ , h ₂ ,..., h _N (N is a positive integer) are established by using the training data D _train ), where each basic predictor h _i (i=1~N) may come from different algorithms, different hyperparameters, or different sampling samples. This case is not limited to this.

給予初始化(t=1，t代表疊代回合數)。例如，將所有樣本x _j (j=1~n)的個別樣本權重

全部都初始化為1，亦即，

=1。經初始化後的處理集合H可表示如後H={h ₁ ,h ₂ ,...,h _N }，處理集合H所包括的預測器是指未被選的預測器的集合，至於選擇預測器的原則將在底下說明之。在本案實施例中，訓練資料D _train 與驗證資料D _valid 同屬於樣本資料。例如但不受限於，目前共有1000筆的樣本資料，其中，800筆樣本資料當成訓練資料，而剩餘的200筆樣本資料則當成驗證資料。當然，本案並不受限於此。 In step 220 (training phase), each sample weight of each sample in the validation data (validation data) D _valid is set (initialized), and the processing set H is set (initialized). The sample weight for each sample x _j (j=1~n, n represents the number of samples in the verification data) in the verification data D _valid

Give initialization (t=1, t represents the number of iteration rounds). For example, the individual sample weights of all samples x _j (j=1~n)

All are initialized to 1, that is,

=1. The initialized processing set H can be expressed as the following H = { h ₁ , h ₂ ,..., h _N }, and the predictors included in the processing set H refer to the set of unselected predictors. As for the selection prediction The principle of the device will be explained below. In the embodiment of this case, the training data D _train and the verification data D _valid belong to the sample data. For example, but not limited to, there are currently 1000 pieces of sample data, of which 800 pieces of sample data are regarded as training data, and the remaining 200 pieces of sample data are regarded as verification data. Of course, this case is not limited to this.

In step 230 (training phase), in the first iteration round, the respective predictors of the predictors in the processing set H are established according to all the sample data of the verification data D _valid and the individual sample weights of each sample data Weighting function.

After each predictor weight function of each predictor is established, each predictor h _i in the processing set H respectively predicts each sample x _j in the verification data D _valid , and distinguishes whether the prediction result is correct or incorrect, And record the forecast results.

For example, but not limited to, starting from the predictor h ₁ , predicting each sample x _j in the verification data D _valid by the predictor h ₁ , and distinguishing whether the prediction result is correct or wrong, all the predictor h ₁ The forecast results are recorded. Next, predict each sample x _j in the verification data D _valid by the predictor h ₂ , and distinguish whether the prediction result is correct or wrong, and record all the prediction results of the predictor h ₂ . Repeat the above process until the last predictor h _n in the processing set H predicts each sample x _j in the verification data D _valid , and distinguish whether the prediction result is correct or wrong, and record all the prediction results of the predictor h _n down.

For example, at the first iteration round (t=1), the prediction result of predictor h ₁ is as follows:

Among them, f ₁ ⁽¹⁾ is high or low, and the judgment is made based on whether the output value of f ₁ ⁽¹⁾ is greater than the preset threshold value (in the embodiment of this case, the output value of f ₁ ⁽¹⁾ follows the sample data. the value of x). Set R1 represents the set of sample data that can be correctly predicted by predictor h ₁ when the output value of f ₁ ⁽¹⁾ is high; similarly, set R2 represents when the output value of f ₁ ⁽¹⁾ is low , the set of sample data that can be correctly predicted by the predictor h ₁ ; the set R3 represents the set of sample data that can be wrongly predicted by the predictor h ₁ when the output value of f ₁ ⁽¹⁾ is high; and the set R4 represents , the set of sample data that can be mispredicted by the predictor h ₁ when the output value of f ₁ ⁽¹⁾ is low. In the embodiment of this case, the reference symbol f _i ^{( t )} represents the predictor weight function trained on the predictor h _i in the t-th iteration round.

In step 240 (training phase), the established predictor weight function is evaluated, and according to the evaluation result, a target predictor weight function is selected from the predictor weight functions established in this iteration round and selected from the processing set A target predictor is selected from H , and the processing set H is updated accordingly (eg, the selected predictor is removed from the processing set H ).

The details of step 240 are, for example, as follows: in each iteration round, when evaluating each predictor weight function of each predictor, the correct score s _i of each predictor weight function is evaluated, and the formula is as follows (of course, this case is not limited to this) , other possible embodiments of this case can use other different fractional formulas, which are all within the scope of the spirit of this case):

代表在第t疊代回合中，用於訓練預測器權重函數的樣本資料x _i 的樣本權重。在所有預測器權重函數中，找出具有最高分數的預測器權重函數，例如，f _i 有最高分數，則在第t疊代回合中，設定h _(t)=h _i 以及g _(t)=f _i ^(t)，並將h _i 由H中移除。例如，在目前疊代回合中，f ₂ 具有最高分數，則在目前疊代回合中，選擇h ₂ 與f ₂ ，且將h ₂ 從H中移除。h _(t)與g _(t)分別代表在第t疊代回合中所選擇的目標預測器與目標預測器權重函數。也就是說，在本案實施例中，預測器權重函數的正確分數代表預測器的預測結果是否一致於預測器權重函數的輸出值，如果分數愈高，則代表預測器的預測結果愈一致於預測器權重函數的輸出值；反之，如果分數愈低，則代表預測器的預測結果愈不一致於預測器權重函數的輸出值。 reference symbol

represents the sample weight of the sample data xi used to train the predictor weight function in the t- _th iteration. Among all predictor weight functions, find the predictor weight function with the highest score, for example, f _i has the highest score, then in the t-th iteration, set h _{( t )} = hi and _g ( _{t )} = f _i ^{( t )} _and remove hi from H. For example, in the current _iteration round, f2 has the highest score, then in the current iteration round, h2 and f2 are _selected , and _h2 is removed from _H. h _{( t )} and g _{( t )} represent the target predictor and target predictor weight function selected in the t-th iteration, respectively. That is to say, in the embodiment of this case, the correct score of the predictor weight function represents whether the prediction result of the predictor is consistent with the output value of the predictor weight function. If the score is higher, it means that the prediction result of the predictor is more consistent with the prediction The output value of the predictor weight function; on the contrary, if the score is lower, it means that the prediction result of the predictor is less consistent with the output value of the predictor weight function.

In step 250 (training phase), the individual sample weights of the sample data in the verification data D _valid are updated, the details of which will be described below. If all iteration loops are not over (step 260 ), the flow returns to step 230 for the next iteration round. Conversely, if all iteration rounds have ended, it means that all target predictor weight functions have been established and trained, and the training phase has ended. An overall predictor can be derived therefrom (step 270).

The overall predictor is obtained according to the established set of basic predictors and their individual predictor weight functions. In the test phase, the test data x is input to the established overall predictor, and the prediction result y can be obtained (step 270). In the embodiment of this case, the obtained prediction result can be displayed on a display (not shown). In addition, in the embodiment of the present case, the test data (for the testing phase), the training data D _train (for the training phase) and the verification data D _valid (for the training phase) are generated, for example, from a signal source, such as But not limited to, it is a sensor. For example, in a smart factory, sensors can be used to sense production parameters such as temperature, humidity, pressure, etc.

In the t-th iteration round, update the sample weight of each sample x _j to k _j ^(t+1) , the update method is as follows, where c >1, α ₃ > α ₄ > α ₂ > α ₁ . The updated sample weights can be used to train the predictor weight function in the (t+1)th iteration round.

=

(亦即調低)。如果該樣本x _j 在第t疊代回合中，被(所選的)預測器h _(t)正確預測且預測器h _(t)的相對應預測器權重函數g _(t)的值為低時，則將該樣本x _j 的樣本權重k _j ^(t+1)更新為

=

(亦即調低)。如果該樣本x _j 在第t疊代回合中，被(所選的)預測 器h _(t)錯誤預測且預測器h _(t)的相對應預測器權重函數g _(t)的值為高時，則將該樣本x _j 的樣本權重k _j ^(t+1)更新為

=

(亦即調高)。如果該樣本x _j 在第t疊代回合中，被(所選的)預測器h _(t)錯誤預測且預測器h _(t)的相對應預測器權重函數g _(t)的值為低時，則將該樣本x _j 的樣本權重k _j ^(t+1)更新為

=

(亦即調高)。 That is, when updating the sample weight of each sample x _j , if the sample x _j is correctly predicted by the (selected) predictor h _{( t )} in the t iteration round and the predictor h _{( t )} When the value of the corresponding predictor weight function g _{( t )} is high, the sample weight k _j ^(t+1) of the sample x _j is updated as

=

(i.e. turn down). If the sample x _j is correctly predicted by the (selected) predictor h _{( t )} in the t iteration round and the value of the corresponding predictor weight function g _{( t )} of the predictor h _{(t )} is low , then the sample weight k _j ^(t+1) of the sample x _j is updated as

=

(i.e. turn down). If the sample x _j is mispredicted by the (selected) predictor h _{( t )} in the t-th iteration and the value of the corresponding predictor weight function g _{( t )} for the predictor h _{(t )} is high , then the sample weight k _j ^(t+1) of the sample x _j is updated as

=

(i.e. turn up). If the sample x _j is mispredicted by the (selected) predictor h _{( t )} in the t-th iteration and the value of the corresponding predictor weight function g _{( t )} for the predictor h _{(t )} is low , then the sample weight k _j ^(t+1) of the sample x _j is updated as

=

(i.e. turn up).

That is to say, in the embodiment of this case, for the samples that are wrongly predicted by the selected predictor h _{( t )} , the weight of the samples in the next round is increased, and for the samples that are correctly predicted by the selected predictor h _{( t )} sample, and lower the weight of the sample in the next round.

That is to say, in the embodiment of the present application, how to update the sample weight of the sample data is determined by evaluating the consistency between the weight function of the predictor and the identification result of the selected predictor. When the consistency between the predictor weight function and the identification result of the selected predictor is high, the sample weight of the sample data is lowered. Conversely, when the consistency between the predictor weight function and the identification result of the selected predictor is low, the sample weight of the sample data is increased. In other words, in the embodiment of this case, the weight adjustment method is adjusted according to "whether the prediction result of the predictor is correct" and "whether the prediction result of the predictor is consistent with the output value of the weight function".

To further illustrate, when the selected predictor's identification result for the sample is correct, and a high predictor weight function is assigned to the predictor, the consistency between the two is high. When the selected predictor's identification result for the sample is wrong, and the predictor is assigned a low predictor weight function, the consistency between the two is high. When the selected predictor's identification result for the sample is correct, and the predictor is assigned a low predictor weight function, the consistency between the two is low. When the selected predictor's identification result for the sample is wrong, and the predictor is assigned a high predictor weight function, the consistency between the two is low.

,

,

,

,

)=(1,1,1,1,1)。 An example will now be given to describe the practice of the embodiment of the present case more clearly. For the convenience of description, in order to carry out 3 iteration rounds (ie t=3), the verification data D _valid includes 5 sample data x ₁ -x ₅ , and let c=1.5, ( α ₁ , α ₂ , α ₃ , α ₄ )=(-1, 0, 2, 1) as an example to illustrate. The initialized processing set H = { h ₁ , h ₂ , h ₃ }. The weight of each sample after initialization is (

,

,

,

,

)=(1,1,1,1,1).

In the first iteration round, each predictor weight function f ₁ ⁽¹⁾ , f ₂ ⁽¹⁾ , f ₃ ⁽¹⁾ of each predictor is established according to each sample data and each sample weight of each sample data. The details of how to establish the weight function may not be particularly limited here.

Calculate the score s for each predictor weight function in the first iteration round. Suppose the identification situation of the predictor h ₁ is as follows:

Then the fraction s ₁ of f ₁ ⁽¹⁾ is as follows:

,

,

,

,

)=(

* 1.5^-1,

* 1.5^-1,

* 1.5¹,

* 1.5⁰,

* 1.5²)。 After calculating the respective scores of each predictor weight function, assuming that in the first iteration round, f ₂ ⁽¹⁾ has the highest score, it means that the predictor weight function f established in the first iteration round ₂ ⁽¹⁾ is optimal, so, in the first iteration round, the target predictor is selected as the predictor h ₂ , and the target predictor weight function is selected as the predictor weight function f ₂ ⁽¹⁾ , that is, h ₍₁₎ = h ₂ , g ₍₁₎ = f ₂ ⁽¹⁾ ( h ₍₁₎ and g ₍₁₎ represent the target predictor and target predictor weight function selected in the first iteration round, respectively ), and update H = { h ₁ , h ₃ } (that is, remove h ₂ from H ), and, according to the results of the first iteration round, the weight of each sample of each sample data (which is used for the second iteration Generation round) is adjusted as follows: (

,

,

,

,

)=(

* 1.5 ^-1 ,

* 1.5 ^-1 ,

* 1.5 ¹ ,

* 1.5 ⁰ ,

* 1.5 ² ).

,

,

,

,

)來建立各預測器的各預測器權重函數f ₁ ⁽²⁾，f ₃ ⁽²⁾(f ₂ ⁽²⁾可以不用再建立)。 In the second iteration round, according to each sample data and each sample weight of each sample data (

,

,

,

,

) to establish each predictor weight function f ₁ ⁽²⁾ and f ₃ ⁽²⁾ of each predictor ( f ₂ ⁽²⁾ may not need to be established again).

Similar to the above case, calculate the predictor weight functions f ₁ ⁽²⁾ , f ₃ ⁽²⁾ of the predictors h ₁ , h ₃ in the processing set H = { h ₁ , h ₃ } in the second iteration The fractions s ₁ and s ₃ . Suppose the identification situation _of the predictor h3 is as follows:

Therefore, the fraction s ₃ of f ₃ ⁽²⁾ is as follows:

,

,

,

,

)=(

* 1.5⁰,

* 1.5¹,

* 1.5²,

* 1.5^-1,

* 1.5^-1)=(

* 1.5^-1 * 1.5⁰,

* 1.5^-1 * 1.5¹,

* 1.5¹ * 1.5²,

* 1.5⁰ * 1.5^-1,

* 1.5² * 1.5^-1)。 After calculating the respective scores of each predictor weight function, assuming that in the second iteration round, f ₃ ⁽²⁾ has the highest score, it means that the predictor weight function f established in the second iteration round ₃ ⁽²⁾ is optimal, so in the second iteration round, the target predictor is selected as h ₃ , and the weight function of the target predictor is selected as f ₃ ⁽²⁾ , that is, h ₍₂₎ = h ₃ , g ₍₂₎ = f ₃ ⁽²⁾ ( h ₍₂₎ and g ₍₂₎₎ represent the target predictor and target predictor weight function selected in the second iteration round, respectively), and update H ={ h ₁ } (that is, removing h ₃ from H ), and, according to the results of the second iteration round, the sample weights of each sample data of the verification data (for the third iteration round) are adjusted as follows: (

,

,

,

,

)=(

* 1.5 ⁰ ,

* 1.5 ¹ ,

* 1.5 ² ,

* 1.5 ^-1 ,

* 1.5 ^-1 )=(

* 1.5 ^-1 * 1.5 ⁰ ,

* 1.5 ^-1 * 1.5 ¹ ,

* 1.5 ¹ * 1.5 ² ,

* 1.5 ⁰ * 1.5 ^-1 ,

* 1.5 ² * 1.5 ^-1 ).

,

,

,

,

)來建立預測器h ₁的預測器權重函數f ₁ ⁽³⁾。所以，在第3疊代回合中，選擇預測器h ₁與預測器權重函數f ₁ ⁽³⁾，亦即，h ₍₃₎=h ₁，g ₍₃₎=f ₁ ⁽³⁾(h ₍₃₎與g ₍₃₎分別代表在第3疊代回合中所選擇出的目標預測器與目標預測器權重函數)，且更新H=

(亦即將h ₁從H移除)。如此，即可以結束整體預測器的建立。 In the third iteration round, according to each sample data and each sample weight of each sample data (

,

,

,

,

) to establish the predictor weight function f ₁ ⁽³⁾ of the predictor h ₁ . Therefore, in the 3rd iteration round, the predictor h ₁ and the predictor weight function f ₁ ⁽³⁾ are selected, that is, h ₍₃₎ = h ₁ , g ₍₃₎ = f ₁ ⁽³⁾ ( h _{( 3)} and g ₍₃₎ respectively represent the target predictor and target predictor weight function selected in the third iteration round), and update H =

(that is, remove h1 from H ₎ . In this way, the establishment of the overall predictor can be completed.

Let there be a test data x _test , and its corresponding prediction result y belongs to two categories (1 or -1), and the values after substituting x _test into h _{( i )} ( x ) and g _{( i )} ( x ) are as follows: ( h ₍₁₎ ( x _test ), h ₍₂₎ ( x _test ), h ₍₃₎ ( x _test ))=(1,-1,-1)

( g ₍₁₎ ( x _test ), g ₍₂₎ ( x _test ), g ₍₃₎ ( x _test ))=(0.3,0.7,0.6)

Then the following ensemble method can be used to get the prediction result:

Method 1: Set the critical value to 0.5, starting from g ₍₁₎ , in order (the order is g ₍₁₎ , g ₍₂₎ , g ₍₃₎ ), select the first output value that can exceed the critical value of the weight function. In the above example, the output value of the weight function g ₍₁₎ ( x _test ) is 0.3, which does not exceed the critical value. Therefore, the weight function g ₍₂₎ is considered next. The output value of the weight function g ₍₂₎ ( x _test ) is 0.7, which exceeds the critical value, so the weight function g ₍₂₎ is chosen. Therefore, the value -1 obtained by the corresponding predictor h ₍₂₎ ( x _test ) of the weight function g ₍₂₎ is used as the overall prediction result of the test data x _test (that is, y = h ₍₂₎ ( x _test ) )=-1).

That is, according to method 1, the function output value can be calculated for each predictor weight function, the first weight function whose output value exceeds the critical value is given the highest weight, and the rest of the weight functions are zero.

Method 2: After normalizing these g _{( i )} , perform a weighted average:

Because the result of the weighted average is less than 0, -1 is taken as the overall prediction result of the test data x _test (y=-1). That is, according to method 2, the function output value of each predictor weight function is normalized and used as the integrated weight of each predictor.

FIG. 3 shows a functional block diagram of an integrated learning and testing system according to another embodiment of the present application. The integrated learning and testing system 300 includes: a basic predictor training model 310 , a predictor weight function training module 320 , an evaluation module 330 and a sample weight adjustment module 340 . The basic predictor training model 310 , the predictor weight function training module 320 , the evaluation module 330 and the sample weight adjustment module 340 can implement the integrated learning and testing method of the above embodiments of this application. The integrated learning test system 300 may further include a display (not shown) for displaying the predicted results. In the embodiment of this case, the integrated learning and testing system 300 is, for example, an electronic device (such as a server) with information processing, and the basic predictor training model 310 , the predictor weight function training module 320 , the evaluation module 330 and the The sample weight adjustment module 340 may be implemented by a hardware circuit with information processing capability (eg, a central processing unit).

In the holistic learning of the embodiment of this case, the weight function is established by considering the diversity of each predictor (prediction model), based on complementarity, with sample weight adjustment, and a weight function training mechanism with sequential optimization . In this way, the embodiment of the present case can effectively find a predictor suitable for individual sample regions, and can dynamically train the weight function to improve the overall prediction effect. Here, the so-called "weight function training mechanism with sequential optimization" means that instead of finding all target weight functions in one iteration, individual target weight functions are found round by round (as described above). Generally, in the t-th iteration round, find g _{( t )} ).

The embodiment of this case can produce special effects on the computer (predicting the result through the target weight function generated round by round), rather than just using the computer as a tool. That is, the embodiment of the present case not only uses a computer, but also uses a specific type of rules to achieve a specific effect of improving the overall learning and prediction system.

The field specifically targeted by the above embodiments of this case is, for example, but not limited to, computer prediction systems. Compared with traditional computer prediction systems, which cannot train an effective predictor weight function in a complex sample space, the embodiments of this case can use sequential The optimized weight function estimation module and steps are used to select the target predictor and the target predictor weight function round by round. Therefore, even in the face of a complex sample space, the embodiment of this case can still effectively train the weight of the predictor function (because in this case, multiple target predictor weight functions are not obtained in one round, but multiple target predictor weight functions are obtained round by round).

To sum up, although the present creation has been disclosed above with embodiments, it is not intended to limit the present creation. Those with ordinary knowledge in the technical field to which this creation belongs can make various changes and modifications without departing from the spirit and scope of this creation. Therefore, the scope of protection of this creation should be determined by the scope of the appended patent application.

300: Holistic Learning Test System

310: Basic Predictor Training Model

320: Predictor weight function training module

330: Evaluation Module

340: Sample weight adjustment module

Claims (7)

An integral learning and prediction system, comprising: a basic predictor training model, which uses a plurality of training data transmitted from a signal source to establish a plurality of basic predictors, wherein the signal source is a sensor; a The predictor weight function training module is connected to the basic predictor training model, initializes a plurality of sample weights of a plurality of sample data, and initializes a processing set. After initialization, the processing set includes corresponding to the basic predictors. a plurality of predictors, and, in a first iteration round, according to the sample data and the sample weights, a plurality of predictor weight functions for the predictors in the processing set are established, and the processing Each of the predictors in the set respectively predicts the sample data to generate a prediction result for each of the sample data; an evaluation module evaluates the predictor weight functions, and obtains a prediction from the sample data according to an evaluation result. selecting a target predictor weight function from the predictor weight functions established in the first iteration round and selecting a target predictor corresponding to the target predictor weight function from the processing set; and a sample weight adjustment module , connecting the evaluation module and the predictor weight function training module, so as to update the processing set and update the sample weights of the sample data, the processing set after the update does not include the target predictor; Wherein, according to the updated sample weights of the sample data, perform the next iteration round to repeat the above operation to select another target predictor from the updated processing set and obtain another target predictor weight function , until all Iterative rounds to select a plurality of target predictor weight functions and a plurality of target predictors to integrate into an overall predictor, wherein the overall predictor includes the target predictor weight functions and the target predictors, and the overall predictor A prediction result of the predictor is further displayed on a display. The system of claim 1, wherein, during initialization, the predictor weight function training module weights the samples of the sample data x _j

Given initialization, j=1~n, n represents the number of the sample data, t represents the number of iteration rounds); the initialized processing set H can be expressed as H = { h ₁ , h ₂ ,... , h _N }, h ₁ , h ₂ , . . . , h _N represent the basic predictors, and the processing set H includes the unselected predictors. The system of claim 2, wherein the evaluation module: in each of the iterative rounds, evaluates an individual score of each of the predictor weight functions, the score representing a corresponding response to the sample data Whether the individual prediction results of the predictor are consistent with the output values of the predictor weight function corresponding to the specific samples, wherein, when the identification result corresponding to the predictor for a specific sample is correct, and for the specific sample If the output value of the predictor weight function of the predictor corresponding to the specific sample is higher than a threshold value, the consistency between the two is high, when the identification result of the predictor corresponding to the specific sample is wrong, and The output value of the predictor weight function of the predictor corresponding to the specific sample is lower than the threshold value, then the consistency between the two is high, when the identification result of the predictor corresponding to the specific sample is correct , and the output value of the predictor weight function of the predictor corresponding to the specific sample is lower than the threshold value, then the consistency between the two is low, when the identification result of the predictor corresponding to the specific sample is wrong, and the output value of the predictor weight function of the predictor corresponding to the specific sample is higher than the threshold value, then the consistency between the two is low; A predictor weight function with the highest score is the target predictor weight function for the iteration round, setting h _{( t )} = hi and g _{( t )} = f _{i (} ^{t )} , h _{( t )} and g _{( t )} respectively represent the weight function of the target predictor and the target predictor selected in the t iteration round. The system of claim 1, wherein when updating the processing set, the selected target predictor is removed from the processing set. The system of claim 4, wherein when the sample weight adjustment module updates the sample weight of the sample data x _j : if the sample data x _j is When the target predictor h _{( t )} predicts correctly and the value of the predictor weight function g _{( t )} of the predictor h _{( t )} is high, the sample weight of the sample data x _j is adjusted down; if the sample When the data x _j is correctly predicted by the target predictor h _{( t )} in the t iteration round and the value of the predictor weight function g _{( t )} of the predictor h _{( t )} is low, then the The sample weight of the sample data x _j is adjusted down; if the sample data x _j is incorrectly predicted by the target predictor h _{( t )} in the t iteration round and the predictor weight of the predictor h _{( t )} is When the value of the function g _{( t )} is high, the sample weight of the sample data x _j is increased; and if the sample data x _j is in the t iteration round, the target predictor h _{( t )} When the prediction is wrong and the value of the predictor weight function g _{( t )} of the predictor h _{( t )} is low, the sample weight of the sample data x _j is adjusted higher. The system of claim 1, wherein, in the overall predictor, individual function output values are calculated for each of the target predictor weight functions, for the target predictor whose function output value exceeds a threshold The weight function is given the highest weight, and the rest of the weight functions are zero. The system of claim 1, wherein, in the overall predictor, the individual function output values of each predictor weight function are normalized to serve as the integrated weight of each of the target predictors.

Images